{
  "schemaVersion": 1,
  "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 1, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 2, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 3, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 4, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 5, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 6, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 7, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    }
  ],
  "cupti_version": 22,
  "cuda_runtime_version": 12040,
  "cuda_driver_version": 12080,
  "distributedInfo": {"backend": "nccl", "rank": 2, "world_size": 8, "pg_count": 1, "pg_config": [{"pg_name": "0", "pg_desc": "default_pg", "backend_config": "cuda:nccl", "pg_size": 8, "ranks": [0, 1, 2, 3, 4, 5, 6, 7]}], "nccl_version": "2.21.5"},
  "record_shapes": 1,
  "trace_id": "767AB2E999714B54BE45A6D6BA18A9D9",
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: DivBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622345443.773, "dur": 111.940,
    "args": {
      "External id": 534529,"Record function id": 0, "Sequence number": 6319630, "Fwd thread id": 1, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "DivBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622345460.621, "dur": 86.247,
    "args": {
      "External id": 534530,"Sequence number": 6319630, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 1
    }
  },
  {
    "ph": "f", "id": 1, "pid": 4183438, "tid": 31367, "ts": 679622345460.621,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 4183438, "tid": 31367,
    "ts": 679622345467.902, "dur": 76.966,
    "args": {
      "External id": 534531,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622345566.182, "dur": 251.598,
    "args": {
      "External id": 534532,"Record function id": 0, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622345618.157, "dur": 119.822,
    "args": {
      "External id": 534533,"Record function id": 0, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.19", "pid": 4183438, "tid": 31367,
    "ts": 679622345646.377, "dur": 79.622,
    "args": {
      "External id": 534534,"Record function id": 0, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622345743.684, "dur": 1.637,
    "args": {
      "External id": 534535,"Sequence number": 6319629, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 6
    }
  },
  {
    "ph": "f", "id": 2, "pid": 4183438, "tid": 31367, "ts": 679622345743.684,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622345749.338, "dur": 63.665,
    "args": {
      "External id": 534536,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622345756.913, "dur": 55.488,
    "args": {
      "External id": 534537,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622345768.025, "dur": 3.445,
    "args": {
      "External id": 534538,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622345827.746, "dur": 16710.958,
    "args": {
      "External id": 534539,"Record function id": 0, "Sequence number": 6319627, "Fwd thread id": 1, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622345829.408, "dur": 16699.044,
    "args": {
      "External id": 534540,"Sequence number": 6319627, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11
    }
  },
  {
    "ph": "f", "id": 3, "pid": 4183438, "tid": 31367, "ts": 679622345829.408,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622345867.730, "dur": 3.713,
    "args": {
      "External id": 534541,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622345876.199, "dur": 16529.843,
    "args": {
      "External id": 534542,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622345878.144, "dur": 16527.518,
    "args": {
      "External id": 534543,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622345883.214, "dur": 7.108,
    "args": {
      "External id": 534544,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622345892.072, "dur": 16511.588,
    "args": {
      "External id": 534545,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 4183438, "tid": 31367,
    "ts": 679622362413.008, "dur": 0.509,
    "args": {
      "External id": 534546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622362415.568, "dur": 3.487,
    "args": {
      "External id": 534547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622362417.464, "dur": 1.349,
    "args": {
      "External id": 534548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 31367,
    "ts": 679622362426.606, "dur": 36.221,
    "args": {
      "External id": 534549,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 4183438, "tid": 31367,
    "ts": 679622362471.481, "dur": 46.447,
    "args": {
      "External id": 534550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 21
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 31367,
    "ts": 679622362473.184, "dur": 44.560,
    "args": {
      "External id": 534551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 22
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 31367,
    "ts": 679622362474.914, "dur": 42.602,
    "args": {
      "External id": 534552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 23
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362552.464, "dur": 19.593,
    "args": {
      "External id": 534553,"Record function id": 0, "Sequence number": 6319626, "Fwd thread id": 1, "Ev Idx": 24
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362554.440, "dur": 14.642,
    "args": {
      "External id": 534554,"Sequence number": 6319626, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 25
    }
  },
  {
    "ph": "f", "id": 4, "pid": 4183438, "tid": 31367, "ts": 679622362554.440,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622362558.516, "dur": 10.334,
    "args": {
      "External id": 534555,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622362562.229, "dur": 6.400,
    "args": {
      "External id": 534556,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 27
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362577.633, "dur": 118.901,
    "args": {
      "External id": 534557,"Record function id": 0, "Sequence number": 6319625, "Fwd thread id": 1, "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362579.223, "dur": 108.185,
    "args": {
      "External id": 534558,"Sequence number": 6319625, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 29
    }
  },
  {
    "ph": "f", "id": 5, "pid": 4183438, "tid": 31367, "ts": 679622362579.223,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622362582.386, "dur": 104.464,
    "args": {
      "External id": 534559,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622362587.011, "dur": 32.099,
    "args": {
      "External id": 534560,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 31
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622362591.374, "dur": 4.659,
    "args": {
      "External id": 534561,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622362597.568, "dur": 21.254,
    "args": {
      "External id": 534562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 33
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622362600.279, "dur": 18.104,
    "args": {
      "External id": 534563,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 34
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622362623.527, "dur": 5.095,
    "args": {
      "External id": 534564,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 35
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622362626.566, "dur": 1.595,
    "args": {
      "External id": 534565,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[4194304, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 36
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622362630.044, "dur": 55.370,
    "args": {
      "External id": 534566,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 37
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362703.269, "dur": 65.747,
    "args": {
      "External id": 534567,"Record function id": 0, "Sequence number": 6319624, "Fwd thread id": 1, "Ev Idx": 38
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362704.410, "dur": 60.896,
    "args": {
      "External id": 534568,"Sequence number": 6319624, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 39
    }
  },
  {
    "ph": "f", "id": 6, "pid": 4183438, "tid": 31367, "ts": 679622362704.410,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622362708.209, "dur": 56.848,
    "args": {
      "External id": 534569,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "2", "3"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 40
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622362711.154, "dur": 25.026,
    "args": {
      "External id": 534570,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 41
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622362714.422, "dur": 3.385,
    "args": {
      "External id": 534571,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 42
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622362718.556, "dur": 17.353,
    "args": {
      "External id": 534572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 43
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622362721.312, "dur": 14.075,
    "args": {
      "External id": 534573,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 44
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622362739.634, "dur": 5.771,
    "args": {
      "External id": 534574,"Record function id": 0, "Concrete Inputs": ["", "2", "3"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 45
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622362743.560, "dur": 1.160,
    "args": {
      "External id": 534575,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "3072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 46
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622362746.392, "dur": 18.076,
    "args": {
      "External id": 534576,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 47
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362773.020, "dur": 99.012,
    "args": {
      "External id": 534577,"Record function id": 0, "Sequence number": 6319623, "Fwd thread id": 1, "Ev Idx": 48
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362773.929, "dur": 94.078,
    "args": {
      "External id": 534578,"Sequence number": 6319623, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 49
    }
  },
  {
    "ph": "f", "id": 7, "pid": 4183438, "tid": 31367, "ts": 679622362773.929,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622362775.723, "dur": 91.987,
    "args": {
      "External id": 534579,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 50
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622362780.945, "dur": 17.902,
    "args": {
      "External id": 534580,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 51
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622362781.859, "dur": 2.306,
    "args": {
      "External id": 534581,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 52
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622362784.918, "dur": 13.686,
    "args": {
      "External id": 534582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 53
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622362785.837, "dur": 12.470,
    "args": {
      "External id": 534583,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 54
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622362801.738, "dur": 2.758,
    "args": {
      "External id": 534584,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 55
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622362803.412, "dur": 0.822,
    "args": {
      "External id": 534585,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 56
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622362805.093, "dur": 61.952,
    "args": {
      "External id": 534586,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 57
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362876.043, "dur": 69.405,
    "args": {
      "External id": 534587,"Record function id": 0, "Sequence number": 6319622, "Fwd thread id": 1, "Ev Idx": 58
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362876.975, "dur": 65.498,
    "args": {
      "External id": 534588,"Sequence number": 6319622, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 59
    }
  },
  {
    "ph": "f", "id": 8, "pid": 4183438, "tid": 31367, "ts": 679622362876.975,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622362878.436, "dur": 63.796,
    "args": {
      "External id": 534589,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 60
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622362879.286, "dur": 19.903,
    "args": {
      "External id": 534590,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 61
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622362882.171, "dur": 3.873,
    "args": {
      "External id": 534591,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 62
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622362886.570, "dur": 12.365,
    "args": {
      "External id": 534592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 63
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622362887.537, "dur": 10.950,
    "args": {
      "External id": 534593,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 64
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622362900.183, "dur": 4.493,
    "args": {
      "External id": 534594,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 65
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622362903.647, "dur": 0.876,
    "args": {
      "External id": 534595,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 66
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622362906.914, "dur": 34.748,
    "args": {
      "External id": 534596,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 67
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362949.371, "dur": 36.096,
    "args": {
      "External id": 534597,"Record function id": 0, "Sequence number": 6319621, "Fwd thread id": 1, "Ev Idx": 68
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622362951.097, "dur": 0.971,
    "args": {
      "External id": 534598,"Sequence number": 6319621, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 69
    }
  },
  {
    "ph": "f", "id": 9, "pid": 4183438, "tid": 31367, "ts": 679622362951.097,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622362954.567, "dur": 28.184,
    "args": {
      "External id": 534599,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 70
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622362956.780, "dur": 25.516,
    "args": {
      "External id": 534600,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 71
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622362963.633, "dur": 0.531,
    "args": {
      "External id": 534601,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 72
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622362990.368, "dur": 1070.724,
    "args": {
      "External id": 534602,"Record function id": 0, "Sequence number": 6319619, "Fwd thread id": 1, "Ev Idx": 73
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622362992.041, "dur": 1037.381,
    "args": {
      "External id": 534603,"Sequence number": 6319619, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 74
    }
  },
  {
    "ph": "f", "id": 10, "pid": 4183438, "tid": 31367, "ts": 679622362992.041,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622363030.844, "dur": 2.439,
    "args": {
      "External id": 534604,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 75
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622363035.791, "dur": 907.165,
    "args": {
      "External id": 534605,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 76
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622363037.453, "dur": 905.224,
    "args": {
      "External id": 534606,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 77
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622363040.226, "dur": 4.180,
    "args": {
      "External id": 534607,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 78
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622363045.231, "dur": 896.304,
    "args": {
      "External id": 534608,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 79
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 4183438, "tid": 31367,
    "ts": 679622363946.021, "dur": 0.283,
    "args": {
      "External id": 534609,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 80
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622363947.505, "dur": 4.372,
    "args": {
      "External id": 534610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 81
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622363950.831, "dur": 0.908,
    "args": {
      "External id": 534611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 82
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 31367,
    "ts": 679622363955.673, "dur": 22.848,
    "args": {
      "External id": 534612,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 83
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 4183438, "tid": 31367,
    "ts": 679622363983.493, "dur": 38.889,
    "args": {
      "External id": 534613,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 84
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 31367,
    "ts": 679622363984.740, "dur": 37.455,
    "args": {
      "External id": 534614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 85
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 31367,
    "ts": 679622363985.732, "dur": 36.154,
    "args": {
      "External id": 534615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 86
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622364037.782, "dur": 19.394,
    "args": {
      "External id": 534616,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 87
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364068.174, "dur": 14.797,
    "args": {
      "External id": 534617,"Record function id": 0, "Sequence number": 6319618, "Fwd thread id": 1, "Ev Idx": 88
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364069.401, "dur": 11.282,
    "args": {
      "External id": 534618,"Sequence number": 6319618, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 89
    }
  },
  {
    "ph": "f", "id": 11, "pid": 4183438, "tid": 31367, "ts": 679622364069.401,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622364074.014, "dur": 6.478,
    "args": {
      "External id": 534619,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 90
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622364076.840, "dur": 3.507,
    "args": {
      "External id": 534620,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 91
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364086.818, "dur": 75.858,
    "args": {
      "External id": 534621,"Record function id": 0, "Sequence number": 6319617, "Fwd thread id": 1, "Ev Idx": 92
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364087.568, "dur": 69.784,
    "args": {
      "External id": 534622,"Sequence number": 6319617, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 93
    }
  },
  {
    "ph": "f", "id": 12, "pid": 4183438, "tid": 31367, "ts": 679622364087.568,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622364089.499, "dur": 67.508,
    "args": {
      "External id": 534623,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 94
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622364091.549, "dur": 36.023,
    "args": {
      "External id": 534624,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 95
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622364094.812, "dur": 2.420,
    "args": {
      "External id": 534625,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 96
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622364108.458, "dur": 18.859,
    "args": {
      "External id": 534626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 97
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622364109.827, "dur": 17.024,
    "args": {
      "External id": 534627,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 98
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622364129.081, "dur": 4.918,
    "args": {
      "External id": 534628,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 99
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622364130.903, "dur": 2.781,
    "args": {
      "External id": 534629,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[4194304, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622364134.984, "dur": 21.255,
    "args": {
      "External id": 534630,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364170.137, "dur": 55.546,
    "args": {
      "External id": 534631,"Record function id": 0, "Sequence number": 6319616, "Fwd thread id": 1, "Ev Idx": 102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364171.282, "dur": 51.372,
    "args": {
      "External id": 534632,"Sequence number": 6319616, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 103
    }
  },
  {
    "ph": "f", "id": 13, "pid": 4183438, "tid": 31367, "ts": 679622364171.282,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622364175.269, "dur": 47.082,
    "args": {
      "External id": 534633,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "2", "2"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622364176.896, "dur": 20.735,
    "args": {
      "External id": 534634,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622364177.767, "dur": 2.856,
    "args": {
      "External id": 534635,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622364181.462, "dur": 15.921,
    "args": {
      "External id": 534636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622364182.212, "dur": 14.777,
    "args": {
      "External id": 534637,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622364198.633, "dur": 6.045,
    "args": {
      "External id": 534638,"Record function id": 0, "Concrete Inputs": ["", "2", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622364203.236, "dur": 0.947,
    "args": {
      "External id": 534639,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622364205.180, "dur": 16.711,
    "args": {
      "External id": 534640,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364231.623, "dur": 91.145,
    "args": {
      "External id": 534641,"Record function id": 0, "Sequence number": 6319615, "Fwd thread id": 1, "Ev Idx": 112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364232.454, "dur": 87.865,
    "args": {
      "External id": 534642,"Sequence number": 6319615, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 113
    }
  },
  {
    "ph": "f", "id": 14, "pid": 4183438, "tid": 31367, "ts": 679622364232.454,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622364235.290, "dur": 84.617,
    "args": {
      "External id": 534643,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622364236.291, "dur": 17.632,
    "args": {
      "External id": 534644,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622364237.277, "dur": 2.377,
    "args": {
      "External id": 534645,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622364240.637, "dur": 13.035,
    "args": {
      "External id": 534646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622364241.427, "dur": 11.958,
    "args": {
      "External id": 534647,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622364254.934, "dur": 3.884,
    "args": {
      "External id": 534648,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622364258.093, "dur": 0.597,
    "args": {
      "External id": 534649,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622364259.620, "dur": 59.589,
    "args": {
      "External id": 534650,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364328.502, "dur": 75.369,
    "args": {
      "External id": 534651,"Record function id": 0, "Sequence number": 6319614, "Fwd thread id": 1, "Ev Idx": 122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364329.570, "dur": 57.993,
    "args": {
      "External id": 534652,"Sequence number": 6319614, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 123
    }
  },
  {
    "ph": "f", "id": 15, "pid": 4183438, "tid": 31367, "ts": 679622364329.570,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622364332.095, "dur": 55.220,
    "args": {
      "External id": 534653,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622364332.822, "dur": 14.562,
    "args": {
      "External id": 534654,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622364333.442, "dur": 1.429,
    "args": {
      "External id": 534655,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622364335.675, "dur": 11.463,
    "args": {
      "External id": 534656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622364336.171, "dur": 10.584,
    "args": {
      "External id": 534657,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622364348.240, "dur": 1.741,
    "args": {
      "External id": 534658,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622364349.331, "dur": 0.503,
    "args": {
      "External id": 534659,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622364350.585, "dur": 36.048,
    "args": {
      "External id": 534660,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622364391.033, "dur": 11.486,
    "args": {
      "External id": 534661,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364407.753, "dur": 33.418,
    "args": {
      "External id": 534662,"Record function id": 0, "Sequence number": 6319613, "Fwd thread id": 1, "Ev Idx": 133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622364410.246, "dur": 0.627,
    "args": {
      "External id": 534663,"Sequence number": 6319613, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 134
    }
  },
  {
    "ph": "f", "id": 16, "pid": 4183438, "tid": 31367, "ts": 679622364410.246,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622364413.060, "dur": 25.901,
    "args": {
      "External id": 534664,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622364414.609, "dur": 23.895,
    "args": {
      "External id": 534665,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622364420.487, "dur": 2.368,
    "args": {
      "External id": 534666,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622364445.440, "dur": 1464.326,
    "args": {
      "External id": 534667,"Record function id": 0, "Sequence number": 6319611, "Fwd thread id": 1, "Ev Idx": 138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622364448.658, "dur": 1437.780,
    "args": {
      "External id": 534668,"Sequence number": 6319611, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 139
    }
  },
  {
    "ph": "f", "id": 17, "pid": 4183438, "tid": 31367, "ts": 679622364448.658,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622364477.449, "dur": 1.944,
    "args": {
      "External id": 534669,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622364481.219, "dur": 1322.582,
    "args": {
      "External id": 534670,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622364482.462, "dur": 1321.083,
    "args": {
      "External id": 534671,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622364484.244, "dur": 4.171,
    "args": {
      "External id": 534672,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622364489.270, "dur": 1313.210,
    "args": {
      "External id": 534673,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 4183438, "tid": 31367,
    "ts": 679622365806.774, "dur": 0.335,
    "args": {
      "External id": 534674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622365808.311, "dur": 2.209,
    "args": {
      "External id": 534675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622365809.476, "dur": 0.926,
    "args": {
      "External id": 534676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 31367,
    "ts": 679622365813.953, "dur": 19.358,
    "args": {
      "External id": 534677,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 4183438, "tid": 31367,
    "ts": 679622365839.124, "dur": 40.587,
    "args": {
      "External id": 534678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 31367,
    "ts": 679622365840.251, "dur": 39.271,
    "args": {
      "External id": 534679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 31367,
    "ts": 679622365843.428, "dur": 35.834,
    "args": {
      "External id": 534680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622365893.430, "dur": 13.599,
    "args": {
      "External id": 534681,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622365916.780, "dur": 10.204,
    "args": {
      "External id": 534682,"Record function id": 0, "Sequence number": 6319610, "Fwd thread id": 1, "Ev Idx": 153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622365918.374, "dur": 6.665,
    "args": {
      "External id": 534683,"Sequence number": 6319610, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 154
    }
  },
  {
    "ph": "f", "id": 18, "pid": 4183438, "tid": 31367, "ts": 679622365918.374,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622365920.383, "dur": 4.447,
    "args": {
      "External id": 534684,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622365921.522, "dur": 3.149,
    "args": {
      "External id": 534685,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622365930.369, "dur": 58.933,
    "args": {
      "External id": 534686,"Record function id": 0, "Sequence number": 6319609, "Fwd thread id": 1, "Ev Idx": 157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622365931.142, "dur": 54.041,
    "args": {
      "External id": 534687,"Sequence number": 6319609, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 158
    }
  },
  {
    "ph": "f", "id": 19, "pid": 4183438, "tid": 31367, "ts": 679622365931.142,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622365932.300, "dur": 52.537,
    "args": {
      "External id": 534688,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622365938.295, "dur": 19.165,
    "args": {
      "External id": 534689,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622365939.502, "dur": 2.421,
    "args": {
      "External id": 534690,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622365942.868, "dur": 14.342,
    "args": {
      "External id": 534691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622365944.305, "dur": 12.445,
    "args": {
      "External id": 534692,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622365958.942, "dur": 4.753,
    "args": {
      "External id": 534693,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622365962.366, "dur": 0.929,
    "args": {
      "External id": 534694,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[4194304, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622365964.566, "dur": 19.538,
    "args": {
      "External id": 534695,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622365992.890, "dur": 51.126,
    "args": {
      "External id": 534696,"Record function id": 0, "Sequence number": 6319608, "Fwd thread id": 1, "Ev Idx": 167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622365993.995, "dur": 47.384,
    "args": {
      "External id": 534697,"Sequence number": 6319608, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 168
    }
  },
  {
    "ph": "f", "id": 20, "pid": 4183438, "tid": 31367, "ts": 679622365993.995,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622365995.291, "dur": 45.799,
    "args": {
      "External id": 534698,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "2", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622365996.931, "dur": 17.777,
    "args": {
      "External id": 534699,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622365999.486, "dur": 2.467,
    "args": {
      "External id": 534700,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622366002.665, "dur": 11.799,
    "args": {
      "External id": 534701,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622366003.227, "dur": 10.899,
    "args": {
      "External id": 534702,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622366015.566, "dur": 8.280,
    "args": {
      "External id": 534703,"Record function id": 0, "Concrete Inputs": ["", "2", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622366020.154, "dur": 3.286,
    "args": {
      "External id": 534704,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "1024"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622366024.823, "dur": 15.690,
    "args": {
      "External id": 534705,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622366047.613, "dur": 77.216,
    "args": {
      "External id": 534706,"Record function id": 0, "Sequence number": 6319607, "Fwd thread id": 1, "Ev Idx": 177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622366048.587, "dur": 72.686,
    "args": {
      "External id": 534707,"Sequence number": 6319607, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 178
    }
  },
  {
    "ph": "f", "id": 21, "pid": 4183438, "tid": 31367, "ts": 679622366048.587,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622366050.140, "dur": 70.758,
    "args": {
      "External id": 534708,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622366051.124, "dur": 17.868,
    "args": {
      "External id": 534709,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622366053.973, "dur": 1.788,
    "args": {
      "External id": 534710,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622366056.383, "dur": 12.344,
    "args": {
      "External id": 534711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622366056.974, "dur": 11.409,
    "args": {
      "External id": 534712,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622366069.770, "dur": 3.589,
    "args": {
      "External id": 534713,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622366072.472, "dur": 0.682,
    "args": {
      "External id": 534714,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622366073.972, "dur": 45.913,
    "args": {
      "External id": 534715,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622366130.669, "dur": 80.396,
    "args": {
      "External id": 534716,"Record function id": 0, "Sequence number": 6319606, "Fwd thread id": 1, "Ev Idx": 187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622366131.641, "dur": 61.652,
    "args": {
      "External id": 534717,"Sequence number": 6319606, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 188
    }
  },
  {
    "ph": "f", "id": 22, "pid": 4183438, "tid": 31367, "ts": 679622366131.641,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622366135.019, "dur": 58.025,
    "args": {
      "External id": 534718,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622366135.951, "dur": 17.351,
    "args": {
      "External id": 534719,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622366138.638, "dur": 2.306,
    "args": {
      "External id": 534720,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622366141.488, "dur": 11.557,
    "args": {
      "External id": 534721,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622366142.310, "dur": 10.421,
    "args": {
      "External id": 534722,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622366154.054, "dur": 1.909,
    "args": {
      "External id": 534723,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622366155.237, "dur": 0.552,
    "args": {
      "External id": 534724,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622366156.441, "dur": 36.004,
    "args": {
      "External id": 534725,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622366196.522, "dur": 13.352,
    "args": {
      "External id": 534726,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622366216.435, "dur": 32.033,
    "args": {
      "External id": 534727,"Record function id": 0, "Sequence number": 6319605, "Fwd thread id": 1, "Ev Idx": 198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622366217.605, "dur": 0.769,
    "args": {
      "External id": 534728,"Sequence number": 6319605, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 199
    }
  },
  {
    "ph": "f", "id": 23, "pid": 4183438, "tid": 31367, "ts": 679622366217.605,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622366219.775, "dur": 23.806,
    "args": {
      "External id": 534729,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622366223.122, "dur": 20.066,
    "args": {
      "External id": 534730,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622366226.959, "dur": 0.783,
    "args": {
      "External id": 534731,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622366252.245, "dur": 1534.252,
    "args": {
      "External id": 534732,"Record function id": 0, "Sequence number": 6319604, "Fwd thread id": 1, "Ev Idx": 203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622366264.272, "dur": 1494.820,
    "args": {
      "External id": 534733,"Sequence number": 6319604, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 204
    }
  },
  {
    "ph": "f", "id": 24, "pid": 4183438, "tid": 31367, "ts": 679622366264.272,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622366289.512, "dur": 3.098,
    "args": {
      "External id": 534734,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622366294.254, "dur": 1339.530,
    "args": {
      "External id": 534735,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622366295.226, "dur": 1338.201,
    "args": {
      "External id": 534736,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622366297.169, "dur": 2.933,
    "args": {
      "External id": 534737,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622366300.842, "dur": 1331.588,
    "args": {
      "External id": 534738,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 4183438, "tid": 31367,
    "ts": 679622367638.373, "dur": 0.330,
    "args": {
      "External id": 534739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622367639.559, "dur": 1.971,
    "args": {
      "External id": 534740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 4183438, "tid": 31367,
    "ts": 679622367640.669, "dur": 0.711,
    "args": {
      "External id": 534741,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 31367,
    "ts": 679622367644.777, "dur": 61.617,
    "args": {
      "External id": 534742,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 4183438, "tid": 31367,
    "ts": 679622367712.205, "dur": 40.481,
    "args": {
      "External id": 534743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 31367,
    "ts": 679622367713.150, "dur": 39.310,
    "args": {
      "External id": 534744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 31367,
    "ts": 679622367714.105, "dur": 37.955,
    "args": {
      "External id": 534745,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622367768.817, "dur": 14.516,
    "args": {
      "External id": 534746,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622367797.013, "dur": 11.631,
    "args": {
      "External id": 534747,"Record function id": 0, "Ev Idx": 218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622367799.263, "dur": 7.684,
    "args": {
      "External id": 534748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622367802.270, "dur": 3.961,
    "args": {
      "External id": 534749,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622367803.048, "dur": 3.071,
    "args": {
      "External id": 534750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367811.920, "dur": 8.621,
    "args": {
      "External id": 534751,"Record function id": 0, "Sequence number": 6319603, "Fwd thread id": 1, "Ev Idx": 222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367812.891, "dur": 5.787,
    "args": {
      "External id": 534752,"Sequence number": 6319603, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 223
    }
  },
  {
    "ph": "f", "id": 25, "pid": 4183438, "tid": 31367, "ts": 679622367812.891,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622367815.024, "dur": 3.455,
    "args": {
      "External id": 534753,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622367815.847, "dur": 2.525,
    "args": {
      "External id": 534754,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367823.931, "dur": 61.342,
    "args": {
      "External id": 534755,"Record function id": 0, "Sequence number": 6319602, "Fwd thread id": 1, "Ev Idx": 226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367824.976, "dur": 56.046,
    "args": {
      "External id": 534756,"Sequence number": 6319602, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 227
    }
  },
  {
    "ph": "f", "id": 26, "pid": 4183438, "tid": 31367, "ts": 679622367824.976,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622367828.738, "dur": 51.954,
    "args": {
      "External id": 534757,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622367832.682, "dur": 20.286,
    "args": {
      "External id": 534758,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622367834.146, "dur": 2.568,
    "args": {
      "External id": 534759,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622367837.760, "dur": 14.915,
    "args": {
      "External id": 534760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622367839.288, "dur": 12.880,
    "args": {
      "External id": 534761,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622367854.324, "dur": 4.680,
    "args": {
      "External id": 534762,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622367857.809, "dur": 0.966,
    "args": {
      "External id": 534763,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[4194304, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622367859.835, "dur": 20.230,
    "args": {
      "External id": 534764,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367889.279, "dur": 51.527,
    "args": {
      "External id": 534765,"Record function id": 0, "Sequence number": 6319601, "Fwd thread id": 1, "Ev Idx": 236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367890.257, "dur": 48.145,
    "args": {
      "External id": 534766,"Sequence number": 6319601, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 237
    }
  },
  {
    "ph": "f", "id": 27, "pid": 4183438, "tid": 31367, "ts": 679622367890.257,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622367891.924, "dur": 46.267,
    "args": {
      "External id": 534767,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "2", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622367893.323, "dur": 20.584,
    "args": {
      "External id": 534768,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622367897.548, "dur": 2.747,
    "args": {
      "External id": 534769,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622367901.121, "dur": 12.453,
    "args": {
      "External id": 534770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622367902.574, "dur": 10.628,
    "args": {
      "External id": 534771,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622367914.842, "dur": 6.123,
    "args": {
      "External id": 534772,"Record function id": 0, "Concrete Inputs": ["", "2", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622367919.662, "dur": 0.853,
    "args": {
      "External id": 534773,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622367921.596, "dur": 16.109,
    "args": {
      "External id": 534774,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367944.187, "dur": 72.908,
    "args": {
      "External id": 534775,"Record function id": 0, "Sequence number": 6319600, "Fwd thread id": 1, "Ev Idx": 246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622367945.080, "dur": 69.044,
    "args": {
      "External id": 534776,"Sequence number": 6319600, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 247
    }
  },
  {
    "ph": "f", "id": 28, "pid": 4183438, "tid": 31367, "ts": 679622367945.080,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622367946.306, "dur": 67.560,
    "args": {
      "External id": 534777,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622367947.159, "dur": 20.180,
    "args": {
      "External id": 534778,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622367951.891, "dur": 2.033,
    "args": {
      "External id": 534779,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622367954.633, "dur": 12.463,
    "args": {
      "External id": 534780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622367955.587, "dur": 11.152,
    "args": {
      "External id": 534781,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622367968.322, "dur": 3.437,
    "args": {
      "External id": 534782,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622367969.551, "dur": 2.048,
    "args": {
      "External id": 534783,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622367972.357, "dur": 40.901,
    "args": {
      "External id": 534784,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622368020.973, "dur": 100.288,
    "args": {
      "External id": 534785,"Record function id": 0, "Sequence number": 6319599, "Fwd thread id": 1, "Ev Idx": 256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622368021.898, "dur": 69.493,
    "args": {
      "External id": 534786,"Sequence number": 6319599, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 257
    }
  },
  {
    "ph": "f", "id": 29, "pid": 4183438, "tid": 31367, "ts": 679622368021.898,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622368023.001, "dur": 68.104,
    "args": {
      "External id": 534787,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], [], []], "Ev Idx": 258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 31367,
    "ts": 679622368024.090, "dur": 24.886,
    "args": {
      "External id": 534788,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622368029.155, "dur": 1.525,
    "args": {
      "External id": 534789,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 1024]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 31367,
    "ts": 679622368034.352, "dur": 14.360,
    "args": {
      "External id": 534790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 31367,
    "ts": 679622368037.321, "dur": 11.069,
    "args": {
      "External id": 534791,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], []], "Ev Idx": 262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622368049.905, "dur": 3.400,
    "args": {
      "External id": 534792,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368052.798, "dur": 0.330,
    "args": {
      "External id": 534793,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622368053.773, "dur": 36.721,
    "args": {
      "External id": 534794,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622368094.001, "dur": 24.644,
    "args": {
      "External id": 534795,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [16777216, 4096, 1024, 1], []], "Input Dims": [[16, 4096, 4, 1024], [16, 4096, 4, 1024], []], "Ev Idx": 266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622368128.097, "dur": 334.855,
    "args": {
      "External id": 534796,"Record function id": 0, "Sequence number": 6319598, "Fwd thread id": 1, "Ev Idx": 267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622368129.693, "dur": 325.245,
    "args": {
      "External id": 534797,"Sequence number": 6319598, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 268
    }
  },
  {
    "ph": "f", "id": 30, "pid": 4183438, "tid": 31367, "ts": 679622368129.693,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622368291.025, "dur": 42.050,
    "args": {
      "External id": 534798,"kernel_hash": "cwcxij4wsirwittcv63raib7filq3solcjt7s5agn3kgmuei6diw", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "262144", "1024", "1", "1986", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/wc/cwcxij4wsirwittcv63raib7filq3solcjt7s5agn3kgmuei6diw.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[262144, 1024], [1024], [262144, 1024], [262144, 1024], [132, 1024], [262144], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_0", "pid": 4183438, "tid": 31367,
    "ts": 679622368362.807, "dur": 26.577,
    "args": {
      "External id": 534799,"kernel_hash": "c56gzotufw6tqupgsmxkhpndljpenudz4jbqivnf244rqggxwo5z", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/56/c56gzotufw6tqupgsmxkhpndljpenudz4jbqivnf244rqggxwo5z.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_1", "pid": 4183438, "tid": 31367,
    "ts": 679622368409.251, "dur": 20.063,
    "args": {
      "External id": 534800,"kernel_hash": "c44qdv7w6al3deyvs2xnwnza27yveqwslv54u6ulmyjvjqyrds6n", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/44/c44qdv7w6al3deyvs2xnwnza27yveqwslv54u6ulmyjvjqyrds6n.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622368471.107, "dur": 12.904,
    "args": {
      "External id": 534801,"Record function id": 0, "Ev Idx": 272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622368473.316, "dur": 9.856,
    "args": {
      "External id": 534802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622368476.100, "dur": 6.214,
    "args": {
      "External id": 534803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622368479.291, "dur": 2.905,
    "args": {
      "External id": 534804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: StackBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622368487.694, "dur": 31.070,
    "args": {
      "External id": 534805,"Record function id": 0, "Sequence number": 6319597, "Fwd thread id": 1, "Ev Idx": 276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "StackBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622368488.969, "dur": 22.525,
    "args": {
      "External id": 534806,"Sequence number": 6319597, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1024, 1]], "Input Dims": [[16, 4096, 4, 1024]], "Ev Idx": 277
    }
  },
  {
    "ph": "f", "id": 31, "pid": 4183438, "tid": 31367, "ts": 679622368488.969,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622368491.012, "dur": 7.250,
    "args": {
      "External id": 534807,"Record function id": 0, "Concrete Inputs": ["", "-2", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368495.487, "dur": 1.018,
    "args": {
      "External id": 534808,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622368499.098, "dur": 4.017,
    "args": {
      "External id": 534809,"Record function id": 0, "Concrete Inputs": ["", "-2", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368502.029, "dur": 0.472,
    "args": {
      "External id": 534810,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "1024"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622368503.635, "dur": 2.040,
    "args": {
      "External id": 534811,"Record function id": 0, "Concrete Inputs": ["", "-2", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368504.654, "dur": 0.504,
    "args": {
      "External id": 534812,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 31367,
    "ts": 679622368506.651, "dur": 4.254,
    "args": {
      "External id": 534813,"Record function id": 0, "Concrete Inputs": ["", "-2", "3"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368509.496, "dur": 0.672,
    "args": {
      "External id": 534814,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "3072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622368522.622, "dur": 5.720,
    "args": {
      "External id": 534815,"Record function id": 0, "Sequence number": 6319596, "Fwd thread id": 1, "Ev Idx": 286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622368523.473, "dur": 1.068,
    "args": {
      "External id": 534816,"Sequence number": 6319596, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 287
    }
  },
  {
    "ph": "f", "id": 32, "pid": 4183438, "tid": 31367, "ts": 679622368523.473,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622368532.482, "dur": 529.615,
    "args": {
      "External id": 534817,"Record function id": 0, "Sequence number": 6319595, "Fwd thread id": 1, "Ev Idx": 288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622368533.399, "dur": 516.204,
    "args": {
      "External id": 534818,"Sequence number": 6319595, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 289
    }
  },
  {
    "ph": "f", "id": 33, "pid": 4183438, "tid": 31367, "ts": 679622368533.399,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622368568.391, "dur": 11.176,
    "args": {
      "External id": 534819,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622368575.296, "dur": 3.907,
    "args": {
      "External id": 534820,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]", "[4096, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[16777216, 4096, 1], [], []], "Input Dims": [[16, 4096, 1024], [], []], "Ev Idx": 291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622368583.083, "dur": 10.033,
    "args": {
      "External id": 534821,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622368585.611, "dur": 6.440,
    "args": {
      "External id": 534822,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368588.915, "dur": 2.881,
    "args": {
      "External id": 534823,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 31367,
    "ts": 679622368596.764, "dur": 163.108,
    "args": {
      "External id": 534824,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4096, 1], [1, 2816], []], "Input Dims": [[65536, 1024], [2816, 1024], []], "Ev Idx": 295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622368598.102, "dur": 3.390,
    "args": {
      "External id": 534825,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2816]], "Input Dims": [[2816, 1024]], "Ev Idx": 296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622368598.990, "dur": 1.674,
    "args": {
      "External id": 534826,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2816], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368600.047, "dur": 0.503,
    "args": {
      "External id": 534827,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[2816, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2816], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 31367,
    "ts": 679622368602.722, "dur": 156.134,
    "args": {
      "External id": 534828,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622368604.656, "dur": 152.484,
    "args": {
      "External id": 534829,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622368768.584, "dur": 7.805,
    "args": {
      "External id": 534830,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [11534336, 2816, 1]], "Input Dims": [[65536, 2816], [16, 4096, 2816]], "Ev Idx": 301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622368772.893, "dur": 3.365,
    "args": {
      "External id": 534831,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622368810.164, "dur": 6.594,
    "args": {
      "External id": 534832,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622368817.946, "dur": 1.913,
    "args": {
      "External id": 534833,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622368820.465, "dur": 1.639,
    "args": {
      "External id": 534834,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622368863.333, "dur": 2.842,
    "args": {
      "External id": 534835,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622368864.654, "dur": 1.374,
    "args": {
      "External id": 534836,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 4183438, "tid": 31367,
    "ts": 679622368892.810, "dur": 135.913,
    "args": {
      "External id": 534837,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[4096, 1], [2816, 1]], []], "Input Dims": [[], [[65536, 1024], [65536, 2816]], []], "Ev Idx": 308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622368898.664, "dur": 7.849,
    "args": {
      "External id": 534838,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368901.759, "dur": 2.466,
    "args": {
      "External id": 534839,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024, 1]", "[4096, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622368911.163, "dur": 5.946,
    "args": {
      "External id": 534840,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4096, 1, 1], []], "Input Dims": [[65536, 1024, 1], []], "Ev Idx": 311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368915.779, "dur": 0.626,
    "args": {
      "External id": 534841,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 65536]", "[1, 1, 4096]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1, 1], [], [], []], "Input Dims": [[65536, 1024, 1], [], [], []], "Ev Idx": 312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622368918.545, "dur": 4.267,
    "args": {
      "External id": 534842,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368921.776, "dur": 0.649,
    "args": {
      "External id": 534843,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622368923.686, "dur": 3.403,
    "args": {
      "External id": 534844,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368925.975, "dur": 0.768,
    "args": {
      "External id": 534845,"Record function id": 0, "Concrete Inputs": ["", "[1, 2816, 65536]", "[1, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1, 1], [], [], []], "Input Dims": [[65536, 2816, 1], [], [], []], "Ev Idx": 316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622368933.348, "dur": 4.340,
    "args": {
      "External id": 534846,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 4096], []], "Input Dims": [[1024, 1, 65536], []], "Ev Idx": 317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368936.772, "dur": 0.609,
    "args": {
      "External id": 534847,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536, 1]", "[1, 4096, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 4096], [], [], []], "Input Dims": [[1024, 1, 65536], [], [], []], "Ev Idx": 318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622368938.474, "dur": 7.398,
    "args": {
      "External id": 534848,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 4096, 1], []], "Input Dims": [[1024, 65536, 1], []], "Ev Idx": 319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622368943.605, "dur": 2.080,
    "args": {
      "External id": 534849,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]", "[1024, 1, 4096]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 4096, 1], [], []], "Input Dims": [[1024, 65536, 1], [], []], "Ev Idx": 320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622368946.811, "dur": 2.163,
    "args": {
      "External id": 534850,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 2816], []], "Input Dims": [[1, 2816, 65536], []], "Ev Idx": 321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622368948.178, "dur": 0.473,
    "args": {
      "External id": 534851,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 2816], [], [], []], "Input Dims": [[1, 2816, 65536], [], [], []], "Ev Idx": 322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622368951.454, "dur": 2.520,
    "args": {
      "External id": 534852,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622368952.297, "dur": 1.578,
    "args": {
      "External id": 534853,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622368956.468, "dur": 60.939,
    "args": {
      "External id": 534854,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1, 4096], [184549376, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816]], "Ev Idx": 325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369018.999, "dur": 1.137,
    "args": {
      "External id": 534855,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2883584, 2816, 1], []], "Input Dims": [[1, 1024, 2816], []], "Ev Idx": 326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622369020.662, "dur": 4.080,
    "args": {
      "External id": 534856,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 2816, 1], []], "Input Dims": [[1024, 1, 2816], []], "Ev Idx": 327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369024.009, "dur": 0.417,
    "args": {
      "External id": 534857,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816, 1]", "[2816, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 2816, 1], [], [], []], "Input Dims": [[1024, 1, 2816], [], [], []], "Ev Idx": 328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369026.666, "dur": 0.867,
    "args": {
      "External id": 534858,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 2816], []], "Input Dims": [[1024, 2816, 1], []], "Ev Idx": 329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369074.575, "dur": 10.168,
    "args": {
      "External id": 534859,"Record function id": 0, "Ev Idx": 330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369076.907, "dur": 7.081,
    "args": {
      "External id": 534860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369079.343, "dur": 3.782,
    "args": {
      "External id": 534861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369080.523, "dur": 2.522,
    "args": {
      "External id": 534862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369088.936, "dur": 25.168,
    "args": {
      "External id": 534863,"Record function id": 0, "Sequence number": 6319594, "Fwd thread id": 1, "Ev Idx": 334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369089.966, "dur": 5.415,
    "args": {
      "External id": 534864,"Sequence number": 6319594, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 335
    }
  },
  {
    "ph": "f", "id": 34, "pid": 4183438, "tid": 31367, "ts": 679622369089.966,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369091.682, "dur": 3.531,
    "args": {
      "External id": 534865,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369094.258, "dur": 0.809,
    "args": {
      "External id": 534866,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369120.684, "dur": 133.243,
    "args": {
      "External id": 534867,"Record function id": 0, "Sequence number": 6319593, "Fwd thread id": 1, "Ev Idx": 338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369123.960, "dur": 122.331,
    "args": {
      "External id": 534868,"Sequence number": 6319593, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 339
    }
  },
  {
    "ph": "f", "id": 35, "pid": 4183438, "tid": 31367, "ts": 679622369123.960,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369128.162, "dur": 5.512,
    "args": {
      "External id": 534869,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369130.344, "dur": 2.514,
    "args": {
      "External id": 534870,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369131.823, "dur": 0.833,
    "args": {
      "External id": 534871,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622369134.958, "dur": 50.284,
    "args": {
      "External id": 534872,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369186.431, "dur": 5.525,
    "args": {
      "External id": 534873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369187.293, "dur": 4.085,
    "args": {
      "External id": 534874,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369190.071, "dur": 1.116,
    "args": {
      "External id": 534875,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369193.360, "dur": 6.916,
    "args": {
      "External id": 534876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369194.191, "dur": 5.645,
    "args": {
      "External id": 534877,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369197.062, "dur": 2.681,
    "args": {
      "External id": 534878,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622369200.847, "dur": 44.466,
    "args": {
      "External id": 534879,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369259.122, "dur": 8.905,
    "args": {
      "External id": 534880,"Record function id": 0, "Sequence number": 6319592, "Fwd thread id": 1, "Ev Idx": 351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369260.107, "dur": 5.833,
    "args": {
      "External id": 534881,"Sequence number": 6319592, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 352
    }
  },
  {
    "ph": "f", "id": 36, "pid": 4183438, "tid": 31367, "ts": 679622369260.107,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369261.508, "dur": 4.259,
    "args": {
      "External id": 534882,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369264.101, "dur": 1.560,
    "args": {
      "External id": 534883,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369271.456, "dur": 7.819,
    "args": {
      "External id": 534884,"Record function id": 0, "Sequence number": 6319591, "Fwd thread id": 1, "Ev Idx": 355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369272.400, "dur": 4.422,
    "args": {
      "External id": 534885,"Sequence number": 6319591, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 356
    }
  },
  {
    "ph": "f", "id": 37, "pid": 4183438, "tid": 31367, "ts": 679622369272.400,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369273.225, "dur": 3.345,
    "args": {
      "External id": 534886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369274.422, "dur": 1.708,
    "args": {
      "External id": 534887,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369275.426, "dur": 0.597,
    "args": {
      "External id": 534888,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369285.219, "dur": 5.732,
    "args": {
      "External id": 534889,"Record function id": 0, "Ev Idx": 360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369286.673, "dur": 3.730,
    "args": {
      "External id": 534890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369287.735, "dur": 2.311,
    "args": {
      "External id": 534891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369288.556, "dur": 1.406,
    "args": {
      "External id": 534892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369296.048, "dur": 7.452,
    "args": {
      "External id": 534893,"Record function id": 0, "Sequence number": 6319590, "Fwd thread id": 1, "Ev Idx": 364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369296.928, "dur": 4.758,
    "args": {
      "External id": 534894,"Sequence number": 6319590, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 365
    }
  },
  {
    "ph": "f", "id": 38, "pid": 4183438, "tid": 31367, "ts": 679622369296.928,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369299.868, "dur": 1.666,
    "args": {
      "External id": 534895,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369300.555, "dur": 0.849,
    "args": {
      "External id": 534896,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369306.780, "dur": 94.724,
    "args": {
      "External id": 534897,"Record function id": 0, "Sequence number": 6319589, "Fwd thread id": 1, "Ev Idx": 368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369307.476, "dur": 86.548,
    "args": {
      "External id": 534898,"Sequence number": 6319589, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 369
    }
  },
  {
    "ph": "f", "id": 39, "pid": 4183438, "tid": 31367, "ts": 679622369307.476,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369309.763, "dur": 4.523,
    "args": {
      "External id": 534899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369310.531, "dur": 3.340,
    "args": {
      "External id": 534900,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369313.279, "dur": 0.489,
    "args": {
      "External id": 534901,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622369314.882, "dur": 28.330,
    "args": {
      "External id": 534902,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369344.186, "dur": 6.890,
    "args": {
      "External id": 534903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369344.827, "dur": 5.730,
    "args": {
      "External id": 534904,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369347.571, "dur": 2.848,
    "args": {
      "External id": 534905,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369352.384, "dur": 3.874,
    "args": {
      "External id": 534906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369352.922, "dur": 2.912,
    "args": {
      "External id": 534907,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369355.267, "dur": 0.500,
    "args": {
      "External id": 534908,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622369356.855, "dur": 36.356,
    "args": {
      "External id": 534909,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369405.848, "dur": 36.344,
    "args": {
      "External id": 534910,"Record function id": 0, "Sequence number": 6319588, "Fwd thread id": 1, "Ev Idx": 381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369406.810, "dur": 4.892,
    "args": {
      "External id": 534911,"Sequence number": 6319588, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 382
    }
  },
  {
    "ph": "f", "id": 40, "pid": 4183438, "tid": 31367, "ts": 679622369406.810,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369407.950, "dur": 3.617,
    "args": {
      "External id": 534912,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369410.246, "dur": 1.172,
    "args": {
      "External id": 534913,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622369415.030, "dur": 24.638,
    "args": {
      "External id": 534914,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369446.242, "dur": 9.208,
    "args": {
      "External id": 534915,"Record function id": 0, "Sequence number": 6319587, "Fwd thread id": 1, "Ev Idx": 386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369447.330, "dur": 6.352,
    "args": {
      "External id": 534916,"Sequence number": 6319587, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 387
    }
  },
  {
    "ph": "f", "id": 41, "pid": 4183438, "tid": 31367, "ts": 679622369447.330,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369449.955, "dur": 3.501,
    "args": {
      "External id": 534917,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369450.882, "dur": 2.101,
    "args": {
      "External id": 534918,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369452.332, "dur": 0.540,
    "args": {
      "External id": 534919,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369459.336, "dur": 4.937,
    "args": {
      "External id": 534920,"Record function id": 0, "Ev Idx": 391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369460.450, "dur": 3.414,
    "args": {
      "External id": 534921,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369461.566, "dur": 1.863,
    "args": {
      "External id": 534922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369462.065, "dur": 1.284,
    "args": {
      "External id": 534923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622369468.299, "dur": 477.917,
    "args": {
      "External id": 534924,"Record function id": 0, "Sequence number": 6319586, "Fwd thread id": 1, "Ev Idx": 395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622369469.805, "dur": 462.263,
    "args": {
      "External id": 534925,"Sequence number": 6319586, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 396
    }
  },
  {
    "ph": "f", "id": 42, "pid": 4183438, "tid": 31367, "ts": 679622369469.805,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 31367,
    "ts": 679622369494.810, "dur": 39.833,
    "args": {
      "External id": 534926,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622369496.282, "dur": 38.067,
    "args": {
      "External id": 534927,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622369500.598, "dur": 6.002,
    "args": {
      "External id": 534928,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622369503.081, "dur": 2.916,
    "args": {
      "External id": 534929,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622369507.947, "dur": 25.797,
    "args": {
      "External id": 534930,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369547.912, "dur": 4.164,
    "args": {
      "External id": 534931,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369549.069, "dur": 2.883,
    "args": {
      "External id": 534932,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369556.294, "dur": 1.601,
    "args": {
      "External id": 534933,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369557.014, "dur": 0.786,
    "args": {
      "External id": 534934,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622369572.126, "dur": 2.148,
    "args": {
      "External id": 534935,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622369588.010, "dur": 2.925,
    "args": {
      "External id": 534936,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369811.912, "dur": 3.954,
    "args": {
      "External id": 534937,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622369820.500, "dur": 38.421,
    "args": {
      "External id": 534938,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369832.521, "dur": 1.030,
    "args": {
      "External id": 534939,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622369865.009, "dur": 31.009,
    "args": {
      "External id": 534940,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622369866.899, "dur": 28.932,
    "args": {
      "External id": 534941,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369871.198, "dur": 3.942,
    "args": {
      "External id": 534942,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622369878.416, "dur": 16.840,
    "args": {
      "External id": 534943,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622369901.349, "dur": 2.570,
    "args": {
      "External id": 534944,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369902.631, "dur": 1.154,
    "args": {
      "External id": 534945,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369910.882, "dur": 6.070,
    "args": {
      "External id": 534946,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369913.314, "dur": 3.525,
    "args": {
      "External id": 534947,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369919.002, "dur": 1.972,
    "args": {
      "External id": 534948,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369920.038, "dur": 0.853,
    "args": {
      "External id": 534949,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369960.177, "dur": 9.063,
    "args": {
      "External id": 534950,"Record function id": 0, "Ev Idx": 421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622369962.066, "dur": 6.404,
    "args": {
      "External id": 534951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369964.426, "dur": 3.167,
    "args": {
      "External id": 534952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622369965.597, "dur": 1.893,
    "args": {
      "External id": 534953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369972.734, "dur": 9.344,
    "args": {
      "External id": 534954,"Record function id": 0, "Sequence number": 6319585, "Fwd thread id": 1, "Ev Idx": 425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369973.769, "dur": 5.680,
    "args": {
      "External id": 534955,"Sequence number": 6319585, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 426
    }
  },
  {
    "ph": "f", "id": 43, "pid": 4183438, "tid": 31367, "ts": 679622369973.769,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622369975.242, "dur": 4.006,
    "args": {
      "External id": 534956,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622369977.829, "dur": 1.308,
    "args": {
      "External id": 534957,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369987.650, "dur": 161.555,
    "args": {
      "External id": 534958,"Record function id": 0, "Sequence number": 6319584, "Fwd thread id": 1, "Ev Idx": 429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622369988.965, "dur": 155.156,
    "args": {
      "External id": 534959,"Sequence number": 6319584, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 430
    }
  },
  {
    "ph": "f", "id": 44, "pid": 4183438, "tid": 31367, "ts": 679622369988.965,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622369992.024, "dur": 4.562,
    "args": {
      "External id": 534960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622369993.570, "dur": 2.394,
    "args": {
      "External id": 534961,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622369994.904, "dur": 0.841,
    "args": {
      "External id": 534962,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622369997.643, "dur": 73.458,
    "args": {
      "External id": 534963,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622370072.396, "dur": 7.695,
    "args": {
      "External id": 534964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622370075.703, "dur": 3.767,
    "args": {
      "External id": 534965,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370077.126, "dur": 2.134,
    "args": {
      "External id": 534966,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622370081.424, "dur": 4.565,
    "args": {
      "External id": 534967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622370082.028, "dur": 3.530,
    "args": {
      "External id": 534968,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370084.857, "dur": 0.625,
    "args": {
      "External id": 534969,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622370086.579, "dur": 56.047,
    "args": {
      "External id": 534970,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370156.206, "dur": 12.467,
    "args": {
      "External id": 534971,"Record function id": 0, "Sequence number": 6319583, "Fwd thread id": 1, "Ev Idx": 442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370157.322, "dur": 9.055,
    "args": {
      "External id": 534972,"Sequence number": 6319583, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 443
    }
  },
  {
    "ph": "f", "id": 45, "pid": 4183438, "tid": 31367, "ts": 679622370157.322,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622370160.919, "dur": 5.305,
    "args": {
      "External id": 534973,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622370162.004, "dur": 4.066,
    "args": {
      "External id": 534974,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370172.272, "dur": 6.113,
    "args": {
      "External id": 534975,"Record function id": 0, "Sequence number": 6319582, "Fwd thread id": 1, "Ev Idx": 446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370173.119, "dur": 3.519,
    "args": {
      "External id": 534976,"Sequence number": 6319582, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 447
    }
  },
  {
    "ph": "f", "id": 46, "pid": 4183438, "tid": 31367, "ts": 679622370173.119,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622370173.746, "dur": 2.666,
    "args": {
      "External id": 534977,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622370174.554, "dur": 1.351,
    "args": {
      "External id": 534978,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370175.388, "dur": 0.415,
    "args": {
      "External id": 534979,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622370184.352, "dur": 7.792,
    "args": {
      "External id": 534980,"Record function id": 0, "Ev Idx": 451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622370185.645, "dur": 5.979,
    "args": {
      "External id": 534981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622370187.160, "dur": 4.118,
    "args": {
      "External id": 534982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622370189.660, "dur": 1.518,
    "args": {
      "External id": 534983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370195.147, "dur": 6.356,
    "args": {
      "External id": 534984,"Record function id": 0, "Sequence number": 6319581, "Fwd thread id": 1, "Ev Idx": 455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370196.050, "dur": 3.451,
    "args": {
      "External id": 534985,"Sequence number": 6319581, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 456
    }
  },
  {
    "ph": "f", "id": 47, "pid": 4183438, "tid": 31367, "ts": 679622370196.050,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622370197.152, "dur": 2.186,
    "args": {
      "External id": 534986,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622370197.945, "dur": 1.251,
    "args": {
      "External id": 534987,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622370207.974, "dur": 300.455,
    "args": {
      "External id": 534988,"Record function id": 0, "Sequence number": 6319580, "Fwd thread id": 1, "Ev Idx": 459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622370209.138, "dur": 283.424,
    "args": {
      "External id": 534989,"Sequence number": 6319580, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 460
    }
  },
  {
    "ph": "f", "id": 48, "pid": 4183438, "tid": 31367, "ts": 679622370209.138,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622370225.783, "dur": 11.167,
    "args": {
      "External id": 534990,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370232.490, "dur": 4.065,
    "args": {
      "External id": 534991,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622370239.222, "dur": 3.180,
    "args": {
      "External id": 534992,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370240.334, "dur": 1.865,
    "args": {
      "External id": 534993,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622370243.917, "dur": 6.064,
    "args": {
      "External id": 534994,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370246.555, "dur": 3.184,
    "args": {
      "External id": 534995,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622370279.607, "dur": 187.109,
    "args": {
      "External id": 534996,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622370373.632, "dur": 4.374,
    "args": {
      "External id": 534997,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622370379.692, "dur": 4.315,
    "args": {
      "External id": 534998,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622370478.755, "dur": 3.501,
    "args": {
      "External id": 534999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622370485.287, "dur": 1.338,
    "args": {
      "External id": 535000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622370488.487, "dur": 1.020,
    "args": {
      "External id": 535001,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622370515.617, "dur": 268.240,
    "args": {
      "External id": 535002,"Record function id": 0, "Sequence number": 6319579, "Fwd thread id": 1, "Ev Idx": 473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622370516.953, "dur": 257.371,
    "args": {
      "External id": 535003,"Sequence number": 6319579, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 474
    }
  },
  {
    "ph": "f", "id": 49, "pid": 4183438, "tid": 31367, "ts": 679622370516.953,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622370536.069, "dur": 41.769,
    "args": {
      "External id": 535004,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370541.009, "dur": 2.667,
    "args": {
      "External id": 535005,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622370544.915, "dur": 32.289,
    "args": {
      "External id": 535006,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622370589.878, "dur": 4.177,
    "args": {
      "External id": 535007,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370591.438, "dur": 2.257,
    "args": {
      "External id": 535008,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622370793.313, "dur": 161.890,
    "args": {
      "External id": 535009,"Record function id": 0, "Sequence number": 6319578, "Fwd thread id": 1, "Ev Idx": 480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622370794.887, "dur": 154.361,
    "args": {
      "External id": 535010,"Sequence number": 6319578, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 481
    }
  },
  {
    "ph": "f", "id": 50, "pid": 4183438, "tid": 31367, "ts": 679622370794.887,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622370807.756, "dur": 32.879,
    "args": {
      "External id": 535011,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370810.888, "dur": 3.718,
    "args": {
      "External id": 535012,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622370815.829, "dur": 24.133,
    "args": {
      "External id": 535013,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622370850.372, "dur": 5.934,
    "args": {
      "External id": 535014,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622370851.835, "dur": 4.121,
    "args": {
      "External id": 535015,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370961.527, "dur": 14.470,
    "args": {
      "External id": 535016,"Record function id": 0, "Sequence number": 6319577, "Fwd thread id": 1, "Ev Idx": 487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370963.025, "dur": 10.614,
    "args": {
      "External id": 535017,"Sequence number": 6319577, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 488
    }
  },
  {
    "ph": "f", "id": 51, "pid": 4183438, "tid": 31367, "ts": 679622370963.025,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622370964.937, "dur": 8.410,
    "args": {
      "External id": 535018,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622370966.499, "dur": 6.673,
    "args": {
      "External id": 535019,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370979.499, "dur": 8.388,
    "args": {
      "External id": 535020,"Record function id": 0, "Sequence number": 6319576, "Fwd thread id": 1, "Ev Idx": 491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370980.559, "dur": 5.468,
    "args": {
      "External id": 535021,"Sequence number": 6319576, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 492
    }
  },
  {
    "ph": "f", "id": 52, "pid": 4183438, "tid": 31367, "ts": 679622370980.559,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622370983.708, "dur": 2.158,
    "args": {
      "External id": 535022,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622370984.492, "dur": 1.238,
    "args": {
      "External id": 535023,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370990.850, "dur": 7.439,
    "args": {
      "External id": 535024,"Record function id": 0, "Sequence number": 6319575, "Fwd thread id": 1, "Ev Idx": 495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622370993.582, "dur": 2.738,
    "args": {
      "External id": 535025,"Sequence number": 6319575, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 496
    }
  },
  {
    "ph": "f", "id": 53, "pid": 4183438, "tid": 31367, "ts": 679622370993.582,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622370994.446, "dur": 1.718,
    "args": {
      "External id": 535026,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622370995.052, "dur": 0.980,
    "args": {
      "External id": 535027,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371001.512, "dur": 7.862,
    "args": {
      "External id": 535028,"Record function id": 0, "Sequence number": 6319574, "Fwd thread id": 1, "Ev Idx": 499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371004.542, "dur": 2.806,
    "args": {
      "External id": 535029,"Sequence number": 6319574, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 500
    }
  },
  {
    "ph": "f", "id": 54, "pid": 4183438, "tid": 31367, "ts": 679622371004.542,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371005.440, "dur": 1.758,
    "args": {
      "External id": 535030,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371006.053, "dur": 1.038,
    "args": {
      "External id": 535031,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371012.745, "dur": 204.848,
    "args": {
      "External id": 535032,"Record function id": 0, "Sequence number": 6319573, "Fwd thread id": 1, "Ev Idx": 503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371013.725, "dur": 193.809,
    "args": {
      "External id": 535033,"Sequence number": 6319573, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 504
    }
  },
  {
    "ph": "f", "id": 55, "pid": 4183438, "tid": 31367, "ts": 679622371013.725,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371018.757, "dur": 6.090,
    "args": {
      "External id": 535034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371020.727, "dur": 3.485,
    "args": {
      "External id": 535035,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371022.411, "dur": 1.523,
    "args": {
      "External id": 535036,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622371026.178, "dur": 97.578,
    "args": {
      "External id": 535037,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371126.181, "dur": 8.094,
    "args": {
      "External id": 535038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371127.799, "dur": 5.406,
    "args": {
      "External id": 535039,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371131.280, "dur": 1.703,
    "args": {
      "External id": 535040,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371156.441, "dur": 4.846,
    "args": {
      "External id": 535041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371157.407, "dur": 3.380,
    "args": {
      "External id": 535042,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371160.314, "dur": 0.358,
    "args": {
      "External id": 535043,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622371162.032, "dur": 44.577,
    "args": {
      "External id": 535044,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371224.660, "dur": 6.478,
    "args": {
      "External id": 535045,"Record function id": 0, "Sequence number": 6319572, "Fwd thread id": 1, "Ev Idx": 516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371225.600, "dur": 3.663,
    "args": {
      "External id": 535046,"Sequence number": 6319572, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 517
    }
  },
  {
    "ph": "f", "id": 56, "pid": 4183438, "tid": 31367, "ts": 679622371225.600,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371226.949, "dur": 2.156,
    "args": {
      "External id": 535047,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371227.545, "dur": 1.408,
    "args": {
      "External id": 535048,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371234.448, "dur": 9.136,
    "args": {
      "External id": 535049,"Record function id": 0, "Sequence number": 6319571, "Fwd thread id": 1, "Ev Idx": 520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371236.966, "dur": 4.939,
    "args": {
      "External id": 535050,"Sequence number": 6319571, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 521
    }
  },
  {
    "ph": "f", "id": 57, "pid": 4183438, "tid": 31367, "ts": 679622371236.966,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371237.583, "dur": 4.080,
    "args": {
      "External id": 535051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371238.080, "dur": 3.095,
    "args": {
      "External id": 535052,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371240.601, "dur": 0.398,
    "args": {
      "External id": 535053,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622371249.016, "dur": 8.958,
    "args": {
      "External id": 535054,"Record function id": 0, "Ev Idx": 525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622371250.319, "dur": 6.755,
    "args": {
      "External id": 535055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622371252.817, "dur": 3.769,
    "args": {
      "External id": 535056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622371253.881, "dur": 2.611,
    "args": {
      "External id": 535057,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371262.882, "dur": 6.399,
    "args": {
      "External id": 535058,"Record function id": 0, "Sequence number": 6319570, "Fwd thread id": 1, "Ev Idx": 529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371263.635, "dur": 3.837,
    "args": {
      "External id": 535059,"Sequence number": 6319570, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 530
    }
  },
  {
    "ph": "f", "id": 58, "pid": 4183438, "tid": 31367, "ts": 679622371263.635,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371264.460, "dur": 2.833,
    "args": {
      "External id": 535060,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371266.504, "dur": 0.697,
    "args": {
      "External id": 535061,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371272.519, "dur": 95.400,
    "args": {
      "External id": 535062,"Record function id": 0, "Sequence number": 6319569, "Fwd thread id": 1, "Ev Idx": 533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371273.266, "dur": 89.078,
    "args": {
      "External id": 535063,"Sequence number": 6319569, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 534
    }
  },
  {
    "ph": "f", "id": 59, "pid": 4183438, "tid": 31367, "ts": 679622371273.266,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371275.086, "dur": 4.143,
    "args": {
      "External id": 535064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371275.532, "dur": 3.251,
    "args": {
      "External id": 535065,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371278.246, "dur": 0.421,
    "args": {
      "External id": 535066,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622371281.837, "dur": 34.216,
    "args": {
      "External id": 535067,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371317.002, "dur": 2.502,
    "args": {
      "External id": 535068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371317.634, "dur": 1.370,
    "args": {
      "External id": 535069,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371318.437, "dur": 0.459,
    "args": {
      "External id": 535070,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371320.660, "dur": 3.959,
    "args": {
      "External id": 535071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371321.170, "dur": 3.046,
    "args": {
      "External id": 535072,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371321.871, "dur": 2.272,
    "args": {
      "External id": 535073,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622371327.044, "dur": 34.512,
    "args": {
      "External id": 535074,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371372.066, "dur": 32.389,
    "args": {
      "External id": 535075,"Record function id": 0, "Sequence number": 6319568, "Fwd thread id": 1, "Ev Idx": 546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371372.856, "dur": 4.417,
    "args": {
      "External id": 535076,"Sequence number": 6319568, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 547
    }
  },
  {
    "ph": "f", "id": 60, "pid": 4183438, "tid": 31367, "ts": 679622371372.856,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371375.415, "dur": 1.706,
    "args": {
      "External id": 535077,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371375.914, "dur": 1.095,
    "args": {
      "External id": 535078,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622371380.237, "dur": 22.123,
    "args": {
      "External id": 535079,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371408.135, "dur": 8.395,
    "args": {
      "External id": 535080,"Record function id": 0, "Sequence number": 6319567, "Fwd thread id": 1, "Ev Idx": 551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371409.055, "dur": 5.520,
    "args": {
      "External id": 535081,"Sequence number": 6319567, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 552
    }
  },
  {
    "ph": "f", "id": 61, "pid": 4183438, "tid": 31367, "ts": 679622371409.055,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371409.755, "dur": 4.613,
    "args": {
      "External id": 535082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371412.198, "dur": 1.666,
    "args": {
      "External id": 535083,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371413.141, "dur": 0.627,
    "args": {
      "External id": 535084,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622371420.296, "dur": 4.408,
    "args": {
      "External id": 535085,"Record function id": 0, "Ev Idx": 556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622371421.319, "dur": 2.976,
    "args": {
      "External id": 535086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622371422.275, "dur": 1.656,
    "args": {
      "External id": 535087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622371422.873, "dur": 0.967,
    "args": {
      "External id": 535088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371428.214, "dur": 6.381,
    "args": {
      "External id": 535089,"Record function id": 0, "Sequence number": 6319566, "Fwd thread id": 1, "Ev Idx": 560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371430.340, "dur": 2.359,
    "args": {
      "External id": 535090,"Sequence number": 6319566, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 561
    }
  },
  {
    "ph": "f", "id": 62, "pid": 4183438, "tid": 31367, "ts": 679622371430.340,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371431.091, "dur": 1.461,
    "args": {
      "External id": 535091,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371431.740, "dur": 0.725,
    "args": {
      "External id": 535092,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371439.640, "dur": 86.687,
    "args": {
      "External id": 535093,"Record function id": 0, "Sequence number": 6319565, "Fwd thread id": 1, "Ev Idx": 564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371440.311, "dur": 80.005,
    "args": {
      "External id": 535094,"Sequence number": 6319565, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 565
    }
  },
  {
    "ph": "f", "id": 63, "pid": 4183438, "tid": 31367, "ts": 679622371440.311,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371441.962, "dur": 2.115,
    "args": {
      "External id": 535095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371442.400, "dur": 1.268,
    "args": {
      "External id": 535096,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371443.126, "dur": 0.442,
    "args": {
      "External id": 535097,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622371444.747, "dur": 34.799,
    "args": {
      "External id": 535098,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371480.613, "dur": 4.630,
    "args": {
      "External id": 535099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371483.143, "dur": 1.610,
    "args": {
      "External id": 535100,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371484.020, "dur": 0.621,
    "args": {
      "External id": 535101,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371486.491, "dur": 3.083,
    "args": {
      "External id": 535102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371486.897, "dur": 2.197,
    "args": {
      "External id": 535103,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371488.616, "dur": 0.385,
    "args": {
      "External id": 535104,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622371489.999, "dur": 29.687,
    "args": {
      "External id": 535105,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371530.528, "dur": 25.553,
    "args": {
      "External id": 535106,"Record function id": 0, "Sequence number": 6319564, "Fwd thread id": 1, "Ev Idx": 577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371533.063, "dur": 3.074,
    "args": {
      "External id": 535107,"Sequence number": 6319564, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 578
    }
  },
  {
    "ph": "f", "id": 64, "pid": 4183438, "tid": 31367, "ts": 679622371533.063,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371533.995, "dur": 1.977,
    "args": {
      "External id": 535108,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371534.706, "dur": 1.119,
    "args": {
      "External id": 535109,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622371538.424, "dur": 15.688,
    "args": {
      "External id": 535110,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371559.551, "dur": 10.714,
    "args": {
      "External id": 535111,"Record function id": 0, "Sequence number": 6319563, "Fwd thread id": 1, "Ev Idx": 582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622371560.341, "dur": 8.215,
    "args": {
      "External id": 535112,"Sequence number": 6319563, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 583
    }
  },
  {
    "ph": "f", "id": 65, "pid": 4183438, "tid": 31367, "ts": 679622371560.341,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622371560.871, "dur": 7.488,
    "args": {
      "External id": 535113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622371562.860, "dur": 5.013,
    "args": {
      "External id": 535114,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371567.095, "dur": 0.676,
    "args": {
      "External id": 535115,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622371574.001, "dur": 4.354,
    "args": {
      "External id": 535116,"Record function id": 0, "Ev Idx": 587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622371575.016, "dur": 2.943,
    "args": {
      "External id": 535117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622371575.762, "dur": 1.693,
    "args": {
      "External id": 535118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622371576.333, "dur": 1.024,
    "args": {
      "External id": 535119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622371582.211, "dur": 419.173,
    "args": {
      "External id": 535120,"Record function id": 0, "Sequence number": 6319562, "Fwd thread id": 1, "Ev Idx": 591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622371583.170, "dur": 383.038,
    "args": {
      "External id": 535121,"Sequence number": 6319562, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 592
    }
  },
  {
    "ph": "f", "id": 66, "pid": 4183438, "tid": 31367, "ts": 679622371583.170,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371614.543, "dur": 1.545,
    "args": {
      "External id": 535122,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371615.025, "dur": 0.933,
    "args": {
      "External id": 535123,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622371630.905, "dur": 5.309,
    "args": {
      "External id": 535124,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622371647.278, "dur": 3.960,
    "args": {
      "External id": 535125,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371848.263, "dur": 3.761,
    "args": {
      "External id": 535126,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622371856.353, "dur": 48.315,
    "args": {
      "External id": 535127,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371868.921, "dur": 1.046,
    "args": {
      "External id": 535128,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622371911.196, "dur": 29.771,
    "args": {
      "External id": 535129,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622371913.012, "dur": 27.719,
    "args": {
      "External id": 535130,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622371917.099, "dur": 3.860,
    "args": {
      "External id": 535131,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622371922.556, "dur": 17.576,
    "args": {
      "External id": 535132,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622371945.508, "dur": 4.667,
    "args": {
      "External id": 535133,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371948.787, "dur": 1.282,
    "args": {
      "External id": 535134,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622371955.824, "dur": 3.230,
    "args": {
      "External id": 535135,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622371958.090, "dur": 0.865,
    "args": {
      "External id": 535136,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622371977.847, "dur": 17.976,
    "args": {
      "External id": 535137,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372015.118, "dur": 9.266,
    "args": {
      "External id": 535138,"Record function id": 0, "Ev Idx": 609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372017.360, "dur": 6.313,
    "args": {
      "External id": 535139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372019.511, "dur": 3.284,
    "args": {
      "External id": 535140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372020.638, "dur": 2.001,
    "args": {
      "External id": 535141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372029.790, "dur": 5.368,
    "args": {
      "External id": 535142,"Record function id": 0, "Sequence number": 6319561, "Fwd thread id": 1, "Ev Idx": 613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372030.935, "dur": 1.251,
    "args": {
      "External id": 535143,"Sequence number": 6319561, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 614
    }
  },
  {
    "ph": "f", "id": 67, "pid": 4183438, "tid": 31367, "ts": 679622372030.935,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622372038.838, "dur": 440.932,
    "args": {
      "External id": 535144,"Record function id": 0, "Sequence number": 6319560, "Fwd thread id": 1, "Ev Idx": 615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622372040.122, "dur": 428.407,
    "args": {
      "External id": 535145,"Sequence number": 6319560, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 616
    }
  },
  {
    "ph": "f", "id": 68, "pid": 4183438, "tid": 31367, "ts": 679622372040.122,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372070.981, "dur": 8.875,
    "args": {
      "External id": 535146,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622372076.386, "dur": 3.193,
    "args": {
      "External id": 535147,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]", "[4096, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[16777216, 4096, 1], [], []], "Input Dims": [[16, 4096, 1024], [], []], "Ev Idx": 618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372082.815, "dur": 8.065,
    "args": {
      "External id": 535148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372084.504, "dur": 5.689,
    "args": {
      "External id": 535149,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372087.152, "dur": 2.867,
    "args": {
      "External id": 535150,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 31367,
    "ts": 679622372096.068, "dur": 122.595,
    "args": {
      "External id": 535151,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4096, 1], [1, 2816], []], "Input Dims": [[65536, 1024], [2816, 1024], []], "Ev Idx": 622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372096.950, "dur": 22.007,
    "args": {
      "External id": 535152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2816]], "Input Dims": [[2816, 1024]], "Ev Idx": 623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372115.686, "dur": 2.449,
    "args": {
      "External id": 535153,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2816], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372117.279, "dur": 0.586,
    "args": {
      "External id": 535154,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[2816, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2816], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 31367,
    "ts": 679622372120.455, "dur": 97.463,
    "args": {
      "External id": 535155,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622372121.731, "dur": 95.345,
    "args": {
      "External id": 535156,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622372223.159, "dur": 5.445,
    "args": {
      "External id": 535157,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [11534336, 2816, 1]], "Input Dims": [[65536, 2816], [16, 4096, 2816]], "Ev Idx": 628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372226.659, "dur": 1.810,
    "args": {
      "External id": 535158,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622372263.457, "dur": 5.042,
    "args": {
      "External id": 535159,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622372269.518, "dur": 2.016,
    "args": {
      "External id": 535160,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622372272.321, "dur": 1.833,
    "args": {
      "External id": 535161,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372308.035, "dur": 2.165,
    "args": {
      "External id": 535162,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372308.686, "dur": 1.342,
    "args": {
      "External id": 535163,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 4183438, "tid": 31367,
    "ts": 679622372333.572, "dur": 115.291,
    "args": {
      "External id": 535164,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[4096, 1], [2816, 1]], []], "Input Dims": [[], [[65536, 1024], [65536, 2816]], []], "Ev Idx": 635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622372338.519, "dur": 7.955,
    "args": {
      "External id": 535165,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372342.993, "dur": 2.695,
    "args": {
      "External id": 535166,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024, 1]", "[4096, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622372347.989, "dur": 6.754,
    "args": {
      "External id": 535167,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4096, 1, 1], []], "Input Dims": [[65536, 1024, 1], []], "Ev Idx": 638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372353.712, "dur": 0.397,
    "args": {
      "External id": 535168,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 65536]", "[1, 1, 4096]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1, 1], [], [], []], "Input Dims": [[65536, 1024, 1], [], [], []], "Ev Idx": 639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622372356.194, "dur": 3.234,
    "args": {
      "External id": 535169,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372358.524, "dur": 0.557,
    "args": {
      "External id": 535170,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622372359.917, "dur": 2.192,
    "args": {
      "External id": 535171,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372360.838, "dur": 0.731,
    "args": {
      "External id": 535172,"Record function id": 0, "Concrete Inputs": ["", "[1, 2816, 65536]", "[1, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1, 1], [], [], []], "Input Dims": [[65536, 2816, 1], [], [], []], "Ev Idx": 643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622372366.788, "dur": 2.033,
    "args": {
      "External id": 535173,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 4096], []], "Input Dims": [[1024, 1, 65536], []], "Ev Idx": 644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372367.971, "dur": 0.571,
    "args": {
      "External id": 535174,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536, 1]", "[1, 4096, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 4096], [], [], []], "Input Dims": [[1024, 1, 65536], [], [], []], "Ev Idx": 645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372372.571, "dur": 5.451,
    "args": {
      "External id": 535175,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 4096, 1], []], "Input Dims": [[1024, 65536, 1], []], "Ev Idx": 646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622372375.676, "dur": 2.128,
    "args": {
      "External id": 535176,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]", "[1024, 1, 4096]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 4096, 1], [], []], "Input Dims": [[1024, 65536, 1], [], []], "Ev Idx": 647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622372378.650, "dur": 2.920,
    "args": {
      "External id": 535177,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 2816], []], "Input Dims": [[1, 2816, 65536], []], "Ev Idx": 648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372380.817, "dur": 0.464,
    "args": {
      "External id": 535178,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 2816], [], [], []], "Input Dims": [[1, 2816, 65536], [], [], []], "Ev Idx": 649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372381.995, "dur": 3.618,
    "args": {
      "External id": 535179,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372383.738, "dur": 1.792,
    "args": {
      "External id": 535180,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622372387.157, "dur": 50.627,
    "args": {
      "External id": 535181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1, 4096], [184549376, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816]], "Ev Idx": 652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372439.310, "dur": 1.166,
    "args": {
      "External id": 535182,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2883584, 2816, 1], []], "Input Dims": [[1, 1024, 2816], []], "Ev Idx": 653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622372441.361, "dur": 3.488,
    "args": {
      "External id": 535183,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 2816, 1], []], "Input Dims": [[1024, 1, 2816], []], "Ev Idx": 654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372443.836, "dur": 0.627,
    "args": {
      "External id": 535184,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816, 1]", "[2816, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 2816, 1], [], [], []], "Input Dims": [[1024, 1, 2816], [], [], []], "Ev Idx": 655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372447.036, "dur": 0.787,
    "args": {
      "External id": 535185,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 2816], []], "Input Dims": [[1024, 2816, 1], []], "Ev Idx": 656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372490.283, "dur": 8.668,
    "args": {
      "External id": 535186,"Record function id": 0, "Ev Idx": 657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372492.316, "dur": 6.046,
    "args": {
      "External id": 535187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372494.208, "dur": 3.360,
    "args": {
      "External id": 535188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372495.305, "dur": 2.179,
    "args": {
      "External id": 535189,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372502.572, "dur": 6.512,
    "args": {
      "External id": 535190,"Record function id": 0, "Sequence number": 6319559, "Fwd thread id": 1, "Ev Idx": 661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372503.528, "dur": 3.444,
    "args": {
      "External id": 535191,"Sequence number": 6319559, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 662
    }
  },
  {
    "ph": "f", "id": 69, "pid": 4183438, "tid": 31367, "ts": 679622372503.528,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372505.071, "dur": 1.692,
    "args": {
      "External id": 535192,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372505.654, "dur": 0.986,
    "args": {
      "External id": 535193,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372512.570, "dur": 104.774,
    "args": {
      "External id": 535194,"Record function id": 0, "Sequence number": 6319558, "Fwd thread id": 1, "Ev Idx": 665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372513.375, "dur": 96.484,
    "args": {
      "External id": 535195,"Sequence number": 6319558, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 666
    }
  },
  {
    "ph": "f", "id": 70, "pid": 4183438, "tid": 31367, "ts": 679622372513.375,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372517.908, "dur": 3.812,
    "args": {
      "External id": 535196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372519.328, "dur": 1.866,
    "args": {
      "External id": 535197,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372520.282, "dur": 0.700,
    "args": {
      "External id": 535198,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622372522.612, "dur": 35.275,
    "args": {
      "External id": 535199,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372558.874, "dur": 7.037,
    "args": {
      "External id": 535200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372559.636, "dur": 5.621,
    "args": {
      "External id": 535201,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372564.074, "dur": 1.000,
    "args": {
      "External id": 535202,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372567.223, "dur": 4.735,
    "args": {
      "External id": 535203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372567.733, "dur": 3.721,
    "args": {
      "External id": 535204,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372568.460, "dur": 2.924,
    "args": {
      "External id": 535205,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622372572.413, "dur": 36.743,
    "args": {
      "External id": 535206,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372621.786, "dur": 6.880,
    "args": {
      "External id": 535207,"Record function id": 0, "Sequence number": 6319557, "Fwd thread id": 1, "Ev Idx": 678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372622.668, "dur": 4.563,
    "args": {
      "External id": 535208,"Sequence number": 6319557, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 679
    }
  },
  {
    "ph": "f", "id": 71, "pid": 4183438, "tid": 31367, "ts": 679622372622.668,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372623.858, "dur": 3.207,
    "args": {
      "External id": 535209,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372625.889, "dur": 1.033,
    "args": {
      "External id": 535210,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372632.191, "dur": 8.162,
    "args": {
      "External id": 535211,"Record function id": 0, "Sequence number": 6319556, "Fwd thread id": 1, "Ev Idx": 682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372634.864, "dur": 3.463,
    "args": {
      "External id": 535212,"Sequence number": 6319556, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 683
    }
  },
  {
    "ph": "f", "id": 72, "pid": 4183438, "tid": 31367, "ts": 679622372634.864,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372635.623, "dur": 2.497,
    "args": {
      "External id": 535213,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372636.218, "dur": 1.436,
    "args": {
      "External id": 535214,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372636.932, "dur": 0.627,
    "args": {
      "External id": 535215,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372643.964, "dur": 4.732,
    "args": {
      "External id": 535216,"Record function id": 0, "Ev Idx": 687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372645.249, "dur": 2.976,
    "args": {
      "External id": 535217,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372646.273, "dur": 1.625,
    "args": {
      "External id": 535218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372646.979, "dur": 0.754,
    "args": {
      "External id": 535219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372651.692, "dur": 51.568,
    "args": {
      "External id": 535220,"Record function id": 0, "Sequence number": 6319555, "Fwd thread id": 1, "Ev Idx": 691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372652.498, "dur": 47.444,
    "args": {
      "External id": 535221,"Sequence number": 6319555, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 692
    }
  },
  {
    "ph": "f", "id": 73, "pid": 4183438, "tid": 31367, "ts": 679622372652.498,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372695.583, "dur": 4.154,
    "args": {
      "External id": 535222,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372697.923, "dur": 1.567,
    "args": {
      "External id": 535223,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372708.192, "dur": 104.351,
    "args": {
      "External id": 535224,"Record function id": 0, "Sequence number": 6319554, "Fwd thread id": 1, "Ev Idx": 695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372709.040, "dur": 94.625,
    "args": {
      "External id": 535225,"Sequence number": 6319554, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 696
    }
  },
  {
    "ph": "f", "id": 74, "pid": 4183438, "tid": 31367, "ts": 679622372709.040,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372711.411, "dur": 5.413,
    "args": {
      "External id": 535226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372712.050, "dur": 4.310,
    "args": {
      "External id": 535227,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372715.404, "dur": 0.839,
    "args": {
      "External id": 535228,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622372719.636, "dur": 35.515,
    "args": {
      "External id": 535229,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372756.466, "dur": 6.217,
    "args": {
      "External id": 535230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372757.131, "dur": 4.978,
    "args": {
      "External id": 535231,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372758.003, "dur": 3.945,
    "args": {
      "External id": 535232,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372763.772, "dur": 5.418,
    "args": {
      "External id": 535233,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372764.440, "dur": 4.248,
    "args": {
      "External id": 535234,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372768.256, "dur": 0.311,
    "args": {
      "External id": 535235,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622372769.624, "dur": 33.307,
    "args": {
      "External id": 535236,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372816.993, "dur": 29.850,
    "args": {
      "External id": 535237,"Record function id": 0, "Sequence number": 6319553, "Fwd thread id": 1, "Ev Idx": 708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372817.865, "dur": 3.249,
    "args": {
      "External id": 535238,"Sequence number": 6319553, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 709
    }
  },
  {
    "ph": "f", "id": 75, "pid": 4183438, "tid": 31367, "ts": 679622372817.865,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372819.064, "dur": 1.905,
    "args": {
      "External id": 535239,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372819.637, "dur": 1.220,
    "args": {
      "External id": 535240,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622372824.115, "dur": 20.173,
    "args": {
      "External id": 535241,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372850.704, "dur": 10.110,
    "args": {
      "External id": 535242,"Record function id": 0, "Sequence number": 6319552, "Fwd thread id": 1, "Ev Idx": 713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622372851.808, "dur": 6.858,
    "args": {
      "External id": 535243,"Sequence number": 6319552, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 714
    }
  },
  {
    "ph": "f", "id": 76, "pid": 4183438, "tid": 31367, "ts": 679622372851.808,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622372852.675, "dur": 5.738,
    "args": {
      "External id": 535244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622372853.528, "dur": 4.364,
    "args": {
      "External id": 535245,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622372857.245, "dur": 0.517,
    "args": {
      "External id": 535246,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372864.831, "dur": 4.593,
    "args": {
      "External id": 535247,"Record function id": 0, "Ev Idx": 718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622372865.866, "dur": 3.082,
    "args": {
      "External id": 535248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372866.770, "dur": 1.851,
    "args": {
      "External id": 535249,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622372867.171, "dur": 1.340,
    "args": {
      "External id": 535250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622372875.304, "dur": 436.610,
    "args": {
      "External id": 535251,"Record function id": 0, "Sequence number": 6319551, "Fwd thread id": 1, "Ev Idx": 722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622372876.306, "dur": 402.262,
    "args": {
      "External id": 535252,"Sequence number": 6319551, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 723
    }
  },
  {
    "ph": "f", "id": 77, "pid": 4183438, "tid": 31367, "ts": 679622372876.306,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 31367,
    "ts": 679622372901.313, "dur": 34.446,
    "args": {
      "External id": 535253,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622372902.712, "dur": 32.843,
    "args": {
      "External id": 535254,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622372905.539, "dur": 5.436,
    "args": {
      "External id": 535255,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622372907.940, "dur": 2.518,
    "args": {
      "External id": 535256,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622372912.582, "dur": 22.350,
    "args": {
      "External id": 535257,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372947.937, "dur": 5.532,
    "args": {
      "External id": 535258,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372950.248, "dur": 3.113,
    "args": {
      "External id": 535259,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622372957.447, "dur": 1.460,
    "args": {
      "External id": 535260,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622372958.038, "dur": 0.777,
    "args": {
      "External id": 535261,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622372971.396, "dur": 2.298,
    "args": {
      "External id": 535262,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622372984.428, "dur": 3.033,
    "args": {
      "External id": 535263,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622373168.900, "dur": 4.018,
    "args": {
      "External id": 535264,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622373177.610, "dur": 33.176,
    "args": {
      "External id": 535265,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373186.924, "dur": 0.849,
    "args": {
      "External id": 535266,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622373216.411, "dur": 29.119,
    "args": {
      "External id": 535267,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622373217.998, "dur": 27.285,
    "args": {
      "External id": 535268,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373222.772, "dur": 4.143,
    "args": {
      "External id": 535269,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622373228.430, "dur": 16.221,
    "args": {
      "External id": 535270,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622373249.939, "dur": 5.152,
    "args": {
      "External id": 535271,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622373253.774, "dur": 1.152,
    "args": {
      "External id": 535272,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622373262.030, "dur": 3.531,
    "args": {
      "External id": 535273,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622373263.169, "dur": 2.282,
    "args": {
      "External id": 535274,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622373267.987, "dur": 1.802,
    "args": {
      "External id": 535275,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622373269.039, "dur": 0.667,
    "args": {
      "External id": 535276,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622373292.824, "dur": 17.569,
    "args": {
      "External id": 535277,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622373325.247, "dur": 10.169,
    "args": {
      "External id": 535278,"Record function id": 0, "Ev Idx": 749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622373327.467, "dur": 7.272,
    "args": {
      "External id": 535279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622373329.492, "dur": 4.338,
    "args": {
      "External id": 535280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622373332.309, "dur": 1.439,
    "args": {
      "External id": 535281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373338.810, "dur": 8.228,
    "args": {
      "External id": 535282,"Record function id": 0, "Sequence number": 6319550, "Fwd thread id": 1, "Ev Idx": 753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373340.025, "dur": 4.370,
    "args": {
      "External id": 535283,"Sequence number": 6319550, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 754
    }
  },
  {
    "ph": "f", "id": 78, "pid": 4183438, "tid": 31367, "ts": 679622373340.025,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622373341.641, "dur": 2.559,
    "args": {
      "External id": 535284,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622373342.616, "dur": 1.430,
    "args": {
      "External id": 535285,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373350.349, "dur": 134.053,
    "args": {
      "External id": 535286,"Record function id": 0, "Sequence number": 6319549, "Fwd thread id": 1, "Ev Idx": 757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373351.125, "dur": 126.884,
    "args": {
      "External id": 535287,"Sequence number": 6319549, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 758
    }
  },
  {
    "ph": "f", "id": 79, "pid": 4183438, "tid": 31367, "ts": 679622373351.125,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622373354.042, "dur": 6.895,
    "args": {
      "External id": 535288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622373357.554, "dur": 2.711,
    "args": {
      "External id": 535289,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373359.247, "dur": 0.794,
    "args": {
      "External id": 535290,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622373362.055, "dur": 66.507,
    "args": {
      "External id": 535291,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622373429.792, "dur": 5.817,
    "args": {
      "External id": 535292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622373430.462, "dur": 4.488,
    "args": {
      "External id": 535293,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373432.125, "dur": 2.644,
    "args": {
      "External id": 535294,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622373439.395, "dur": 3.515,
    "args": {
      "External id": 535295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622373440.593, "dur": 1.830,
    "args": {
      "External id": 535296,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373441.995, "dur": 0.310,
    "args": {
      "External id": 535297,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622373443.359, "dur": 33.717,
    "args": {
      "External id": 535298,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373488.973, "dur": 8.442,
    "args": {
      "External id": 535299,"Record function id": 0, "Sequence number": 6319548, "Fwd thread id": 1, "Ev Idx": 770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373489.763, "dur": 5.588,
    "args": {
      "External id": 535300,"Sequence number": 6319548, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 771
    }
  },
  {
    "ph": "f", "id": 80, "pid": 4183438, "tid": 31367, "ts": 679622373489.763,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622373491.229, "dur": 3.980,
    "args": {
      "External id": 535301,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622373492.417, "dur": 2.682,
    "args": {
      "External id": 535302,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373502.444, "dur": 8.725,
    "args": {
      "External id": 535303,"Record function id": 0, "Sequence number": 6319547, "Fwd thread id": 1, "Ev Idx": 774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373505.408, "dur": 4.282,
    "args": {
      "External id": 535304,"Sequence number": 6319547, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 775
    }
  },
  {
    "ph": "f", "id": 81, "pid": 4183438, "tid": 31367, "ts": 679622373505.408,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622373506.017, "dur": 3.455,
    "args": {
      "External id": 535305,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622373506.683, "dur": 2.262,
    "args": {
      "External id": 535306,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373508.329, "dur": 0.497,
    "args": {
      "External id": 535307,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622373515.082, "dur": 4.866,
    "args": {
      "External id": 535308,"Record function id": 0, "Ev Idx": 779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622373516.775, "dur": 2.649,
    "args": {
      "External id": 535309,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622373517.593, "dur": 1.575,
    "args": {
      "External id": 535310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622373518.082, "dur": 0.994,
    "args": {
      "External id": 535311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373522.915, "dur": 8.452,
    "args": {
      "External id": 535312,"Record function id": 0, "Sequence number": 6319546, "Fwd thread id": 1, "Ev Idx": 783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622373523.709, "dur": 5.964,
    "args": {
      "External id": 535313,"Sequence number": 6319546, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 784
    }
  },
  {
    "ph": "f", "id": 82, "pid": 4183438, "tid": 31367, "ts": 679622373523.709,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622373527.345, "dur": 2.174,
    "args": {
      "External id": 535314,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622373528.148, "dur": 1.236,
    "args": {
      "External id": 535315,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622373535.383, "dur": 323.833,
    "args": {
      "External id": 535316,"Record function id": 0, "Sequence number": 6319545, "Fwd thread id": 1, "Ev Idx": 787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622373536.802, "dur": 301.449,
    "args": {
      "External id": 535317,"Sequence number": 6319545, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 788
    }
  },
  {
    "ph": "f", "id": 83, "pid": 4183438, "tid": 31367, "ts": 679622373536.802,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622373551.873, "dur": 6.244,
    "args": {
      "External id": 535318,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373554.687, "dur": 3.050,
    "args": {
      "External id": 535319,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622373560.277, "dur": 4.315,
    "args": {
      "External id": 535320,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373562.042, "dur": 2.282,
    "args": {
      "External id": 535321,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622373566.075, "dur": 4.255,
    "args": {
      "External id": 535322,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373567.602, "dur": 2.530,
    "args": {
      "External id": 535323,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622373593.691, "dur": 218.029,
    "args": {
      "External id": 535324,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622373715.982, "dur": 5.016,
    "args": {
      "External id": 535325,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622373723.194, "dur": 4.346,
    "args": {
      "External id": 535326,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622373824.536, "dur": 4.263,
    "args": {
      "External id": 535327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622373831.876, "dur": 0.632,
    "args": {
      "External id": 535328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622373834.597, "dur": 0.905,
    "args": {
      "External id": 535329,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622373869.197, "dur": 223.186,
    "args": {
      "External id": 535330,"Record function id": 0, "Sequence number": 6319544, "Fwd thread id": 1, "Ev Idx": 801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622373871.509, "dur": 213.986,
    "args": {
      "External id": 535331,"Sequence number": 6319544, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 802
    }
  },
  {
    "ph": "f", "id": 84, "pid": 4183438, "tid": 31367, "ts": 679622373871.509,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622373891.558, "dur": 47.281,
    "args": {
      "External id": 535332,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373895.292, "dur": 3.125,
    "args": {
      "External id": 535333,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622373899.941, "dur": 38.215,
    "args": {
      "External id": 535334,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622373949.102, "dur": 4.304,
    "args": {
      "External id": 535335,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622373950.907, "dur": 2.176,
    "args": {
      "External id": 535336,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622374115.562, "dur": 157.853,
    "args": {
      "External id": 535337,"Record function id": 0, "Sequence number": 6319543, "Fwd thread id": 1, "Ev Idx": 808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622374118.459, "dur": 148.129,
    "args": {
      "External id": 535338,"Sequence number": 6319543, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 809
    }
  },
  {
    "ph": "f", "id": 85, "pid": 4183438, "tid": 31367, "ts": 679622374118.459,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622374132.873, "dur": 32.688,
    "args": {
      "External id": 535339,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374135.627, "dur": 3.613,
    "args": {
      "External id": 535340,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622374140.402, "dur": 24.615,
    "args": {
      "External id": 535341,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622374173.394, "dur": 4.778,
    "args": {
      "External id": 535342,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374175.633, "dur": 2.222,
    "args": {
      "External id": 535343,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374281.606, "dur": 16.372,
    "args": {
      "External id": 535344,"Record function id": 0, "Sequence number": 6319542, "Fwd thread id": 1, "Ev Idx": 815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374283.040, "dur": 12.250,
    "args": {
      "External id": 535345,"Sequence number": 6319542, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 816
    }
  },
  {
    "ph": "f", "id": 86, "pid": 4183438, "tid": 31367, "ts": 679622374283.040,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374285.637, "dur": 9.362,
    "args": {
      "External id": 535346,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374288.951, "dur": 5.813,
    "args": {
      "External id": 535347,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374301.689, "dur": 6.682,
    "args": {
      "External id": 535348,"Record function id": 0, "Sequence number": 6319541, "Fwd thread id": 1, "Ev Idx": 819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374302.728, "dur": 3.636,
    "args": {
      "External id": 535349,"Sequence number": 6319541, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 820
    }
  },
  {
    "ph": "f", "id": 87, "pid": 4183438, "tid": 31367, "ts": 679622374302.728,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374303.922, "dur": 2.274,
    "args": {
      "External id": 535350,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374305.071, "dur": 1.002,
    "args": {
      "External id": 535351,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374313.295, "dur": 5.663,
    "args": {
      "External id": 535352,"Record function id": 0, "Sequence number": 6319540, "Fwd thread id": 1, "Ev Idx": 823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374314.003, "dur": 3.375,
    "args": {
      "External id": 535353,"Sequence number": 6319540, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 824
    }
  },
  {
    "ph": "f", "id": 88, "pid": 4183438, "tid": 31367, "ts": 679622374314.003,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374315.146, "dur": 2.078,
    "args": {
      "External id": 535354,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374316.261, "dur": 0.821,
    "args": {
      "External id": 535355,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374322.744, "dur": 8.230,
    "args": {
      "External id": 535356,"Record function id": 0, "Sequence number": 6319539, "Fwd thread id": 1, "Ev Idx": 827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374323.867, "dur": 5.167,
    "args": {
      "External id": 535357,"Sequence number": 6319539, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 828
    }
  },
  {
    "ph": "f", "id": 89, "pid": 4183438, "tid": 31367, "ts": 679622374323.867,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374324.720, "dur": 4.165,
    "args": {
      "External id": 535358,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374327.863, "dur": 0.928,
    "args": {
      "External id": 535359,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374334.326, "dur": 157.667,
    "args": {
      "External id": 535360,"Record function id": 0, "Sequence number": 6319538, "Fwd thread id": 1, "Ev Idx": 831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374335.026, "dur": 149.494,
    "args": {
      "External id": 535361,"Sequence number": 6319538, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 832
    }
  },
  {
    "ph": "f", "id": 90, "pid": 4183438, "tid": 31367, "ts": 679622374335.026,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374338.841, "dur": 7.260,
    "args": {
      "External id": 535362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374340.986, "dur": 4.486,
    "args": {
      "External id": 535363,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374343.591, "dur": 1.592,
    "args": {
      "External id": 535364,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622374349.066, "dur": 82.204,
    "args": {
      "External id": 535365,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374432.632, "dur": 5.316,
    "args": {
      "External id": 535366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374433.636, "dur": 3.627,
    "args": {
      "External id": 535367,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374435.636, "dur": 1.376,
    "args": {
      "External id": 535368,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374439.338, "dur": 4.709,
    "args": {
      "External id": 535369,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374440.098, "dur": 3.433,
    "args": {
      "External id": 535370,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374443.036, "dur": 0.428,
    "args": {
      "External id": 535371,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622374444.773, "dur": 38.960,
    "args": {
      "External id": 535372,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374496.693, "dur": 7.202,
    "args": {
      "External id": 535373,"Record function id": 0, "Sequence number": 6319537, "Fwd thread id": 1, "Ev Idx": 844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374497.725, "dur": 4.586,
    "args": {
      "External id": 535374,"Sequence number": 6319537, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 845
    }
  },
  {
    "ph": "f", "id": 91, "pid": 4183438, "tid": 31367, "ts": 679622374497.725,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374499.430, "dur": 2.729,
    "args": {
      "External id": 535375,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374500.544, "dur": 1.471,
    "args": {
      "External id": 535376,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374507.219, "dur": 9.109,
    "args": {
      "External id": 535377,"Record function id": 0, "Sequence number": 6319536, "Fwd thread id": 1, "Ev Idx": 848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374508.127, "dur": 6.090,
    "args": {
      "External id": 535378,"Sequence number": 6319536, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 849
    }
  },
  {
    "ph": "f", "id": 92, "pid": 4183438, "tid": 31367, "ts": 679622374508.127,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374508.866, "dur": 5.154,
    "args": {
      "External id": 535379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374509.481, "dur": 4.038,
    "args": {
      "External id": 535380,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374512.870, "dur": 0.567,
    "args": {
      "External id": 535381,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622374522.186, "dur": 10.019,
    "args": {
      "External id": 535382,"Record function id": 0, "Ev Idx": 853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622374524.165, "dur": 7.234,
    "args": {
      "External id": 535383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622374526.920, "dur": 4.033,
    "args": {
      "External id": 535384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622374528.221, "dur": 2.633,
    "args": {
      "External id": 535385,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374535.409, "dur": 6.035,
    "args": {
      "External id": 535386,"Record function id": 0, "Sequence number": 6319535, "Fwd thread id": 1, "Ev Idx": 857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374536.648, "dur": 3.290,
    "args": {
      "External id": 535387,"Sequence number": 6319535, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 858
    }
  },
  {
    "ph": "f", "id": 93, "pid": 4183438, "tid": 31367, "ts": 679622374536.648,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374537.934, "dur": 1.836,
    "args": {
      "External id": 535388,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374538.845, "dur": 0.813,
    "args": {
      "External id": 535389,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374544.651, "dur": 92.936,
    "args": {
      "External id": 535390,"Record function id": 0, "Sequence number": 6319534, "Fwd thread id": 1, "Ev Idx": 861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374545.447, "dur": 86.869,
    "args": {
      "External id": 535391,"Sequence number": 6319534, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 862
    }
  },
  {
    "ph": "f", "id": 94, "pid": 4183438, "tid": 31367, "ts": 679622374545.447,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374549.701, "dur": 3.848,
    "args": {
      "External id": 535392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374550.505, "dur": 2.612,
    "args": {
      "External id": 535393,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374552.358, "dur": 0.662,
    "args": {
      "External id": 535394,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622374554.171, "dur": 32.793,
    "args": {
      "External id": 535395,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374587.911, "dur": 4.366,
    "args": {
      "External id": 535396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374588.783, "dur": 2.834,
    "args": {
      "External id": 535397,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374590.524, "dur": 0.969,
    "args": {
      "External id": 535398,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374595.783, "dur": 4.455,
    "args": {
      "External id": 535399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374596.374, "dur": 3.322,
    "args": {
      "External id": 535400,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374597.551, "dur": 2.060,
    "args": {
      "External id": 535401,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622374600.794, "dur": 30.848,
    "args": {
      "External id": 535402,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374641.956, "dur": 74.786,
    "args": {
      "External id": 535403,"Record function id": 0, "Sequence number": 6319533, "Fwd thread id": 1, "Ev Idx": 874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374642.729, "dur": 4.060,
    "args": {
      "External id": 535404,"Sequence number": 6319533, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 875
    }
  },
  {
    "ph": "f", "id": 95, "pid": 4183438, "tid": 31367, "ts": 679622374642.729,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374644.312, "dur": 2.333,
    "args": {
      "External id": 535405,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374645.389, "dur": 1.106,
    "args": {
      "External id": 535406,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622374649.773, "dur": 62.335,
    "args": {
      "External id": 535407,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374722.695, "dur": 10.112,
    "args": {
      "External id": 535408,"Record function id": 0, "Sequence number": 6319532, "Fwd thread id": 1, "Ev Idx": 879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374724.147, "dur": 6.209,
    "args": {
      "External id": 535409,"Sequence number": 6319532, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 880
    }
  },
  {
    "ph": "f", "id": 96, "pid": 4183438, "tid": 31367, "ts": 679622374724.147,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374725.166, "dur": 4.965,
    "args": {
      "External id": 535410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374726.646, "dur": 2.741,
    "args": {
      "External id": 535411,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374728.686, "dur": 0.572,
    "args": {
      "External id": 535412,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622374737.406, "dur": 8.754,
    "args": {
      "External id": 535413,"Record function id": 0, "Ev Idx": 884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622374739.274, "dur": 6.344,
    "args": {
      "External id": 535414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622374741.011, "dur": 4.282,
    "args": {
      "External id": 535415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622374743.701, "dur": 1.472,
    "args": {
      "External id": 535416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374749.672, "dur": 6.510,
    "args": {
      "External id": 535417,"Record function id": 0, "Sequence number": 6319531, "Fwd thread id": 1, "Ev Idx": 888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374751.542, "dur": 2.953,
    "args": {
      "External id": 535418,"Sequence number": 6319531, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 889
    }
  },
  {
    "ph": "f", "id": 97, "pid": 4183438, "tid": 31367, "ts": 679622374751.542,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374752.639, "dur": 1.713,
    "args": {
      "External id": 535419,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374753.227, "dur": 1.021,
    "args": {
      "External id": 535420,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374760.985, "dur": 105.244,
    "args": {
      "External id": 535421,"Record function id": 0, "Sequence number": 6319530, "Fwd thread id": 1, "Ev Idx": 892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374761.886, "dur": 96.545,
    "args": {
      "External id": 535422,"Sequence number": 6319530, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 893
    }
  },
  {
    "ph": "f", "id": 98, "pid": 4183438, "tid": 31367, "ts": 679622374761.886,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374763.970, "dur": 4.504,
    "args": {
      "External id": 535423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374764.458, "dur": 3.504,
    "args": {
      "External id": 535424,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374767.315, "dur": 0.540,
    "args": {
      "External id": 535425,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622374769.077, "dur": 41.663,
    "args": {
      "External id": 535426,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374812.040, "dur": 4.872,
    "args": {
      "External id": 535427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374812.678, "dur": 3.600,
    "args": {
      "External id": 535428,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374814.594, "dur": 1.548,
    "args": {
      "External id": 535429,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374817.878, "dur": 8.515,
    "args": {
      "External id": 535430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374823.837, "dur": 2.141,
    "args": {
      "External id": 535431,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374825.278, "dur": 0.633,
    "args": {
      "External id": 535432,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622374827.413, "dur": 30.311,
    "args": {
      "External id": 535433,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374870.663, "dur": 24.593,
    "args": {
      "External id": 535434,"Record function id": 0, "Sequence number": 6319529, "Fwd thread id": 1, "Ev Idx": 905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374871.788, "dur": 3.996,
    "args": {
      "External id": 535435,"Sequence number": 6319529, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 906
    }
  },
  {
    "ph": "f", "id": 99, "pid": 4183438, "tid": 31367, "ts": 679622374871.788,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374873.189, "dur": 2.444,
    "args": {
      "External id": 535436,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374874.295, "dur": 1.228,
    "args": {
      "External id": 535437,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622374878.271, "dur": 15.033,
    "args": {
      "External id": 535438,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374898.812, "dur": 9.241,
    "args": {
      "External id": 535439,"Record function id": 0, "Sequence number": 6319528, "Fwd thread id": 1, "Ev Idx": 910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622374901.440, "dur": 4.795,
    "args": {
      "External id": 535440,"Sequence number": 6319528, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 911
    }
  },
  {
    "ph": "f", "id": 100, "pid": 4183438, "tid": 31367, "ts": 679622374901.440,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622374902.274, "dur": 3.763,
    "args": {
      "External id": 535441,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622374903.084, "dur": 2.479,
    "args": {
      "External id": 535442,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622374905.113, "dur": 0.317,
    "args": {
      "External id": 535443,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622374912.168, "dur": 4.975,
    "args": {
      "External id": 535444,"Record function id": 0, "Ev Idx": 915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622374913.808, "dur": 2.892,
    "args": {
      "External id": 535445,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622374914.635, "dur": 1.781,
    "args": {
      "External id": 535446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622374915.218, "dur": 1.055,
    "args": {
      "External id": 535447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622374921.562, "dur": 383.541,
    "args": {
      "External id": 535448,"Record function id": 0, "Sequence number": 6319527, "Fwd thread id": 1, "Ev Idx": 919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622374923.303, "dur": 351.870,
    "args": {
      "External id": 535449,"Sequence number": 6319527, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 920
    }
  },
  {
    "ph": "f", "id": 101, "pid": 4183438, "tid": 31367, "ts": 679622374923.303,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622374960.160, "dur": 3.817,
    "args": {
      "External id": 535450,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622374962.808, "dur": 1.022,
    "args": {
      "External id": 535451,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622374979.635, "dur": 5.800,
    "args": {
      "External id": 535452,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622374994.776, "dur": 2.549,
    "args": {
      "External id": 535453,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375166.152, "dur": 2.898,
    "args": {
      "External id": 535454,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622375173.471, "dur": 36.869,
    "args": {
      "External id": 535455,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375185.614, "dur": 0.862,
    "args": {
      "External id": 535456,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622375219.972, "dur": 32.457,
    "args": {
      "External id": 535457,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622375221.619, "dur": 30.564,
    "args": {
      "External id": 535458,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375228.758, "dur": 5.348,
    "args": {
      "External id": 535459,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622375235.816, "dur": 15.797,
    "args": {
      "External id": 535460,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622375257.090, "dur": 2.619,
    "args": {
      "External id": 535461,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375258.519, "dur": 1.048,
    "args": {
      "External id": 535462,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375265.563, "dur": 2.208,
    "args": {
      "External id": 535463,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375266.721, "dur": 0.953,
    "args": {
      "External id": 535464,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622375287.000, "dur": 14.605,
    "args": {
      "External id": 535465,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622375317.013, "dur": 8.842,
    "args": {
      "External id": 535466,"Record function id": 0, "Ev Idx": 937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622375319.437, "dur": 5.710,
    "args": {
      "External id": 535467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622375321.286, "dur": 2.794,
    "args": {
      "External id": 535468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622375322.358, "dur": 1.615,
    "args": {
      "External id": 535469,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375329.123, "dur": 5.758,
    "args": {
      "External id": 535470,"Record function id": 0, "Sequence number": 6319526, "Fwd thread id": 1, "Ev Idx": 941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375330.141, "dur": 1.232,
    "args": {
      "External id": 535471,"Sequence number": 6319526, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 942
    }
  },
  {
    "ph": "f", "id": 102, "pid": 4183438, "tid": 31367, "ts": 679622375330.141,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622375338.847, "dur": 432.487,
    "args": {
      "External id": 535472,"Record function id": 0, "Sequence number": 6319525, "Fwd thread id": 1, "Ev Idx": 943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622375340.085, "dur": 419.452,
    "args": {
      "External id": 535473,"Sequence number": 6319525, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 944
    }
  },
  {
    "ph": "f", "id": 103, "pid": 4183438, "tid": 31367, "ts": 679622375340.085,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375370.651, "dur": 9.623,
    "args": {
      "External id": 535474,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622375376.927, "dur": 3.085,
    "args": {
      "External id": 535475,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]", "[4096, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[16777216, 4096, 1], [], []], "Input Dims": [[16, 4096, 1024], [], []], "Ev Idx": 946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622375383.345, "dur": 4.819,
    "args": {
      "External id": 535476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622375384.484, "dur": 2.926,
    "args": {
      "External id": 535477,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375386.722, "dur": 0.539,
    "args": {
      "External id": 535478,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 31367,
    "ts": 679622375391.627, "dur": 81.535,
    "args": {
      "External id": 535479,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4096, 1], [1, 2816], []], "Input Dims": [[65536, 1024], [2816, 1024], []], "Ev Idx": 950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622375392.565, "dur": 6.759,
    "args": {
      "External id": 535480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2816]], "Input Dims": [[2816, 1024]], "Ev Idx": 951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622375395.247, "dur": 3.584,
    "args": {
      "External id": 535481,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2816], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375396.336, "dur": 2.420,
    "args": {
      "External id": 535482,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[2816, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2816], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 31367,
    "ts": 679622375400.531, "dur": 72.095,
    "args": {
      "External id": 535483,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622375401.795, "dur": 70.040,
    "args": {
      "External id": 535484,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622375476.873, "dur": 3.265,
    "args": {
      "External id": 535485,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [11534336, 2816, 1]], "Input Dims": [[65536, 2816], [16, 4096, 2816]], "Ev Idx": 956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375478.576, "dur": 1.406,
    "args": {
      "External id": 535486,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622375512.942, "dur": 4.318,
    "args": {
      "External id": 535487,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622375518.802, "dur": 1.673,
    "args": {
      "External id": 535488,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622375521.551, "dur": 1.538,
    "args": {
      "External id": 535489,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375553.587, "dur": 2.348,
    "args": {
      "External id": 535490,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375554.665, "dur": 1.127,
    "args": {
      "External id": 535491,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 4183438, "tid": 31367,
    "ts": 679622375579.324, "dur": 158.051,
    "args": {
      "External id": 535492,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[4096, 1], [2816, 1]], []], "Input Dims": [[], [[65536, 1024], [65536, 2816]], []], "Ev Idx": 963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622375583.958, "dur": 4.973,
    "args": {
      "External id": 535493,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375587.415, "dur": 0.753,
    "args": {
      "External id": 535494,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024, 1]", "[4096, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622375590.866, "dur": 8.367,
    "args": {
      "External id": 535495,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4096, 1, 1], []], "Input Dims": [[65536, 1024, 1], []], "Ev Idx": 966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375596.122, "dur": 2.498,
    "args": {
      "External id": 535496,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 65536]", "[1, 1, 4096]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1, 1], [], [], []], "Input Dims": [[65536, 1024, 1], [], [], []], "Ev Idx": 967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622375600.407, "dur": 2.277,
    "args": {
      "External id": 535497,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375602.008, "dur": 0.344,
    "args": {
      "External id": 535498,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622375603.559, "dur": 2.820,
    "args": {
      "External id": 535499,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375605.567, "dur": 0.419,
    "args": {
      "External id": 535500,"Record function id": 0, "Concrete Inputs": ["", "[1, 2816, 65536]", "[1, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1, 1], [], [], []], "Input Dims": [[65536, 2816, 1], [], [], []], "Ev Idx": 971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622375610.199, "dur": 3.121,
    "args": {
      "External id": 535501,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 4096], []], "Input Dims": [[1024, 1, 65536], []], "Ev Idx": 972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375612.379, "dur": 0.634,
    "args": {
      "External id": 535502,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536, 1]", "[1, 4096, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 4096], [], [], []], "Input Dims": [[1024, 1, 65536], [], [], []], "Ev Idx": 973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375614.337, "dur": 5.507,
    "args": {
      "External id": 535503,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 4096, 1], []], "Input Dims": [[1024, 65536, 1], []], "Ev Idx": 974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622375617.882, "dur": 1.791,
    "args": {
      "External id": 535504,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]", "[1024, 1, 4096]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 4096, 1], [], []], "Input Dims": [[1024, 65536, 1], [], []], "Ev Idx": 975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622375620.905, "dur": 2.750,
    "args": {
      "External id": 535505,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 2816], []], "Input Dims": [[1, 2816, 65536], []], "Ev Idx": 976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375622.894, "dur": 0.417,
    "args": {
      "External id": 535506,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 2816], [], [], []], "Input Dims": [[1, 2816, 65536], [], [], []], "Ev Idx": 977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375626.635, "dur": 3.573,
    "args": {
      "External id": 535507,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375627.520, "dur": 2.588,
    "args": {
      "External id": 535508,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622375631.470, "dur": 87.676,
    "args": {
      "External id": 535509,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1, 4096], [184549376, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816]], "Ev Idx": 980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375722.643, "dur": 4.172,
    "args": {
      "External id": 535510,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2883584, 2816, 1], []], "Input Dims": [[1, 1024, 2816], []], "Ev Idx": 981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622375728.310, "dur": 3.945,
    "args": {
      "External id": 535511,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 2816, 1], []], "Input Dims": [[1024, 1, 2816], []], "Ev Idx": 982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375730.903, "dur": 0.822,
    "args": {
      "External id": 535512,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816, 1]", "[2816, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 2816, 1], [], [], []], "Input Dims": [[1024, 1, 2816], [], [], []], "Ev Idx": 983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375735.392, "dur": 0.932,
    "args": {
      "External id": 535513,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 2816], []], "Input Dims": [[1024, 2816, 1], []], "Ev Idx": 984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622375783.252, "dur": 9.663,
    "args": {
      "External id": 535514,"Record function id": 0, "Ev Idx": 985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622375785.808, "dur": 6.427,
    "args": {
      "External id": 535515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622375787.961, "dur": 3.430,
    "args": {
      "External id": 535516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622375789.217, "dur": 2.046,
    "args": {
      "External id": 535517,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375796.581, "dur": 8.031,
    "args": {
      "External id": 535518,"Record function id": 0, "Sequence number": 6319524, "Fwd thread id": 1, "Ev Idx": 989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375797.941, "dur": 4.332,
    "args": {
      "External id": 535519,"Sequence number": 6319524, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 990
    }
  },
  {
    "ph": "f", "id": 104, "pid": 4183438, "tid": 31367, "ts": 679622375797.941,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375799.654, "dur": 2.424,
    "args": {
      "External id": 535520,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375800.763, "dur": 1.179,
    "args": {
      "External id": 535521,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375807.987, "dur": 109.409,
    "args": {
      "External id": 535522,"Record function id": 0, "Sequence number": 6319523, "Fwd thread id": 1, "Ev Idx": 993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375808.974, "dur": 102.247,
    "args": {
      "External id": 535523,"Sequence number": 6319523, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 994
    }
  },
  {
    "ph": "f", "id": 105, "pid": 4183438, "tid": 31367, "ts": 679622375808.974,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622375812.416, "dur": 5.865,
    "args": {
      "External id": 535524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622375813.808, "dur": 3.939,
    "args": {
      "External id": 535525,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375817.041, "dur": 0.525,
    "args": {
      "External id": 535526,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622375819.289, "dur": 39.366,
    "args": {
      "External id": 535527,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622375859.751, "dur": 4.630,
    "args": {
      "External id": 535528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622375860.620, "dur": 3.174,
    "args": {
      "External id": 535529,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 1000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375862.472, "dur": 1.160,
    "args": {
      "External id": 535530,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 1001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622375865.565, "dur": 5.372,
    "args": {
      "External id": 535531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622375867.932, "dur": 2.573,
    "args": {
      "External id": 535532,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375869.911, "dur": 0.495,
    "args": {
      "External id": 535533,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622375871.559, "dur": 38.750,
    "args": {
      "External id": 535534,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 1005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375921.773, "dur": 7.014,
    "args": {
      "External id": 535535,"Record function id": 0, "Sequence number": 6319522, "Fwd thread id": 1, "Ev Idx": 1006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375922.870, "dur": 4.579,
    "args": {
      "External id": 535536,"Sequence number": 6319522, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1007
    }
  },
  {
    "ph": "f", "id": 106, "pid": 4183438, "tid": 31367, "ts": 679622375922.870,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375924.591, "dur": 2.700,
    "args": {
      "External id": 535537,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375925.941, "dur": 1.200,
    "args": {
      "External id": 535538,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375932.145, "dur": 8.551,
    "args": {
      "External id": 535539,"Record function id": 0, "Sequence number": 6319521, "Fwd thread id": 1, "Ev Idx": 1010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375932.806, "dur": 5.920,
    "args": {
      "External id": 535540,"Sequence number": 6319521, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1011
    }
  },
  {
    "ph": "f", "id": 107, "pid": 4183438, "tid": 31367, "ts": 679622375932.806,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622375933.661, "dur": 4.841,
    "args": {
      "External id": 535541,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622375935.697, "dur": 2.333,
    "args": {
      "External id": 535542,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375937.454, "dur": 0.429,
    "args": {
      "External id": 535543,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622375944.731, "dur": 5.476,
    "args": {
      "External id": 535544,"Record function id": 0, "Ev Idx": 1015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622375946.286, "dur": 3.468,
    "args": {
      "External id": 535545,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622375947.452, "dur": 2.037,
    "args": {
      "External id": 535546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622375948.014, "dur": 1.355,
    "args": {
      "External id": 535547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375955.089, "dur": 6.268,
    "args": {
      "External id": 535548,"Record function id": 0, "Sequence number": 6319520, "Fwd thread id": 1, "Ev Idx": 1019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375956.289, "dur": 3.180,
    "args": {
      "External id": 535549,"Sequence number": 6319520, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 1020
    }
  },
  {
    "ph": "f", "id": 108, "pid": 4183438, "tid": 31367, "ts": 679622375956.289,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622375957.674, "dur": 1.640,
    "args": {
      "External id": 535550,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622375958.510, "dur": 0.650,
    "args": {
      "External id": 535551,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375966.203, "dur": 107.033,
    "args": {
      "External id": 535552,"Record function id": 0, "Sequence number": 6319519, "Fwd thread id": 1, "Ev Idx": 1023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622375986.437, "dur": 80.091,
    "args": {
      "External id": 535553,"Sequence number": 6319519, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 1024
    }
  },
  {
    "ph": "f", "id": 109, "pid": 4183438, "tid": 31367, "ts": 679622375986.437,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622375988.796, "dur": 2.711,
    "args": {
      "External id": 535554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 1025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622375989.222, "dur": 1.814,
    "args": {
      "External id": 535555,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 1026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622375990.587, "dur": 0.335,
    "args": {
      "External id": 535556,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 1027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622375992.067, "dur": 27.378,
    "args": {
      "External id": 535557,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 1028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622376022.257, "dur": 5.164,
    "args": {
      "External id": 535558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622376022.722, "dur": 4.073,
    "args": {
      "External id": 535559,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 1030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376024.622, "dur": 1.995,
    "args": {
      "External id": 535560,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 1031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622376028.453, "dur": 4.576,
    "args": {
      "External id": 535561,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622376029.202, "dur": 3.405,
    "args": {
      "External id": 535562,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376030.375, "dur": 2.088,
    "args": {
      "External id": 535563,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622376035.043, "dur": 30.805,
    "args": {
      "External id": 535564,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 1035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376077.382, "dur": 55.589,
    "args": {
      "External id": 535565,"Record function id": 0, "Sequence number": 6319518, "Fwd thread id": 1, "Ev Idx": 1036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376078.222, "dur": 3.790,
    "args": {
      "External id": 535566,"Sequence number": 6319518, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1037
    }
  },
  {
    "ph": "f", "id": 110, "pid": 4183438, "tid": 31367, "ts": 679622376078.222,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376079.845, "dur": 2.007,
    "args": {
      "External id": 535567,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376080.745, "dur": 0.979,
    "args": {
      "External id": 535568,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622376085.136, "dur": 44.266,
    "args": {
      "External id": 535569,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376138.751, "dur": 9.622,
    "args": {
      "External id": 535570,"Record function id": 0, "Sequence number": 6319517, "Fwd thread id": 1, "Ev Idx": 1041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376140.238, "dur": 6.405,
    "args": {
      "External id": 535571,"Sequence number": 6319517, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1042
    }
  },
  {
    "ph": "f", "id": 111, "pid": 4183438, "tid": 31367, "ts": 679622376140.238,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622376141.076, "dur": 5.340,
    "args": {
      "External id": 535572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622376142.048, "dur": 3.737,
    "args": {
      "External id": 535573,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376145.087, "dur": 0.550,
    "args": {
      "External id": 535574,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622376152.363, "dur": 5.656,
    "args": {
      "External id": 535575,"Record function id": 0, "Ev Idx": 1046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622376153.782, "dur": 3.724,
    "args": {
      "External id": 535576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622376155.060, "dur": 1.996,
    "args": {
      "External id": 535577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622376155.502, "dur": 1.461,
    "args": {
      "External id": 535578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622376162.003, "dur": 401.316,
    "args": {
      "External id": 535579,"Record function id": 0, "Sequence number": 6319516, "Fwd thread id": 1, "Ev Idx": 1050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622376163.626, "dur": 368.437,
    "args": {
      "External id": 535580,"Sequence number": 6319516, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 1051
    }
  },
  {
    "ph": "f", "id": 112, "pid": 4183438, "tid": 31367, "ts": 679622376163.626,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 31367,
    "ts": 679622376189.100, "dur": 35.231,
    "args": {
      "External id": 535581,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622376190.343, "dur": 33.792,
    "args": {
      "External id": 535582,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622376193.160, "dur": 7.100,
    "args": {
      "External id": 535583,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 1054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622376196.162, "dur": 3.652,
    "args": {
      "External id": 535584,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622376201.645, "dur": 21.767,
    "args": {
      "External id": 535585,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376236.673, "dur": 2.671,
    "args": {
      "External id": 535586,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376237.937, "dur": 1.283,
    "args": {
      "External id": 535587,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376245.560, "dur": 3.242,
    "args": {
      "External id": 535588,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376246.679, "dur": 2.038,
    "args": {
      "External id": 535589,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622376261.153, "dur": 2.197,
    "args": {
      "External id": 535590,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622376274.589, "dur": 2.741,
    "args": {
      "External id": 535591,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376428.359, "dur": 2.510,
    "args": {
      "External id": 535592,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 1063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622376435.047, "dur": 30.185,
    "args": {
      "External id": 535593,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 1064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376443.189, "dur": 0.668,
    "args": {
      "External id": 535594,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 1065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622376470.968, "dur": 28.017,
    "args": {
      "External id": 535595,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 1066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622376472.859, "dur": 25.876,
    "args": {
      "External id": 535596,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 1067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376477.144, "dur": 4.118,
    "args": {
      "External id": 535597,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622376482.524, "dur": 15.588,
    "args": {
      "External id": 535598,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 1069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622376503.556, "dur": 3.000,
    "args": {
      "External id": 535599,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 1070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376505.179, "dur": 1.236,
    "args": {
      "External id": 535600,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 1071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376514.840, "dur": 2.197,
    "args": {
      "External id": 535601,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376515.648, "dur": 1.275,
    "args": {
      "External id": 535602,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376519.188, "dur": 4.132,
    "args": {
      "External id": 535603,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376520.225, "dur": 2.995,
    "args": {
      "External id": 535604,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622376544.131, "dur": 17.585,
    "args": {
      "External id": 535605,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622376572.835, "dur": 9.700,
    "args": {
      "External id": 535606,"Record function id": 0, "Ev Idx": 1077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622376575.489, "dur": 6.509,
    "args": {
      "External id": 535607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622376576.822, "dur": 4.440,
    "args": {
      "External id": 535608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622376579.693, "dur": 1.464,
    "args": {
      "External id": 535609,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376586.350, "dur": 7.992,
    "args": {
      "External id": 535610,"Record function id": 0, "Sequence number": 6319515, "Fwd thread id": 1, "Ev Idx": 1081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376587.648, "dur": 4.509,
    "args": {
      "External id": 535611,"Sequence number": 6319515, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1082
    }
  },
  {
    "ph": "f", "id": 113, "pid": 4183438, "tid": 31367, "ts": 679622376587.648,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376589.423, "dur": 2.540,
    "args": {
      "External id": 535612,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376590.477, "dur": 1.339,
    "args": {
      "External id": 535613,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376598.033, "dur": 188.583,
    "args": {
      "External id": 535614,"Record function id": 0, "Sequence number": 6319514, "Fwd thread id": 1, "Ev Idx": 1085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376598.898, "dur": 179.675,
    "args": {
      "External id": 535615,"Sequence number": 6319514, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1086
    }
  },
  {
    "ph": "f", "id": 114, "pid": 4183438, "tid": 31367, "ts": 679622376598.898,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622376602.020, "dur": 5.976,
    "args": {
      "External id": 535616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622376602.879, "dur": 4.420,
    "args": {
      "External id": 535617,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376606.158, "dur": 0.882,
    "args": {
      "External id": 535618,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622376609.025, "dur": 110.787,
    "args": {
      "External id": 535619,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622376722.741, "dur": 7.288,
    "args": {
      "External id": 535620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622376723.969, "dur": 5.133,
    "args": {
      "External id": 535621,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376725.952, "dur": 2.955,
    "args": {
      "External id": 535622,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622376731.388, "dur": 6.026,
    "args": {
      "External id": 535623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622376734.303, "dur": 2.610,
    "args": {
      "External id": 535624,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376736.141, "dur": 0.702,
    "args": {
      "External id": 535625,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622376738.096, "dur": 39.728,
    "args": {
      "External id": 535626,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376793.510, "dur": 9.864,
    "args": {
      "External id": 535627,"Record function id": 0, "Sequence number": 6319513, "Fwd thread id": 1, "Ev Idx": 1098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376794.707, "dur": 7.072,
    "args": {
      "External id": 535628,"Sequence number": 6319513, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1099
    }
  },
  {
    "ph": "f", "id": 115, "pid": 4183438, "tid": 31367, "ts": 679622376794.707,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376796.351, "dur": 5.267,
    "args": {
      "External id": 535629,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376797.941, "dur": 3.535,
    "args": {
      "External id": 535630,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376806.708, "dur": 9.230,
    "args": {
      "External id": 535631,"Record function id": 0, "Sequence number": 6319512, "Fwd thread id": 1, "Ev Idx": 1102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376807.440, "dur": 6.519,
    "args": {
      "External id": 535632,"Sequence number": 6319512, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1103
    }
  },
  {
    "ph": "f", "id": 116, "pid": 4183438, "tid": 31367, "ts": 679622376807.440,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622376808.594, "dur": 5.167,
    "args": {
      "External id": 535633,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622376810.892, "dur": 2.367,
    "args": {
      "External id": 535634,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376812.492, "dur": 0.653,
    "args": {
      "External id": 535635,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622376820.256, "dur": 6.301,
    "args": {
      "External id": 535636,"Record function id": 0, "Ev Idx": 1107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622376821.944, "dur": 4.075,
    "args": {
      "External id": 535637,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622376823.300, "dur": 2.450,
    "args": {
      "External id": 535638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622376824.019, "dur": 1.623,
    "args": {
      "External id": 535639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376831.194, "dur": 6.180,
    "args": {
      "External id": 535640,"Record function id": 0, "Sequence number": 6319511, "Fwd thread id": 1, "Ev Idx": 1111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622376832.009, "dur": 3.566,
    "args": {
      "External id": 535641,"Sequence number": 6319511, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1112
    }
  },
  {
    "ph": "f", "id": 117, "pid": 4183438, "tid": 31367, "ts": 679622376832.009,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622376833.163, "dur": 2.245,
    "args": {
      "External id": 535642,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622376833.965, "dur": 1.302,
    "args": {
      "External id": 535643,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622376841.587, "dur": 352.831,
    "args": {
      "External id": 535644,"Record function id": 0, "Sequence number": 6319510, "Fwd thread id": 1, "Ev Idx": 1115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622376842.979, "dur": 334.875,
    "args": {
      "External id": 535645,"Sequence number": 6319510, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1116
    }
  },
  {
    "ph": "f", "id": 118, "pid": 4183438, "tid": 31367, "ts": 679622376842.979,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622376858.536, "dur": 7.812,
    "args": {
      "External id": 535646,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376861.708, "dur": 4.286,
    "args": {
      "External id": 535647,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622376868.240, "dur": 4.241,
    "args": {
      "External id": 535648,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376869.679, "dur": 2.614,
    "args": {
      "External id": 535649,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622376882.334, "dur": 5.159,
    "args": {
      "External id": 535650,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622376885.714, "dur": 1.554,
    "args": {
      "External id": 535651,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622376915.891, "dur": 233.417,
    "args": {
      "External id": 535652,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622376993.480, "dur": 5.272,
    "args": {
      "External id": 535653,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622377000.517, "dur": 4.443,
    "args": {
      "External id": 535654,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622377164.108, "dur": 4.557,
    "args": {
      "External id": 535655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622377171.566, "dur": 0.892,
    "args": {
      "External id": 535656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622377174.363, "dur": 0.544,
    "args": {
      "External id": 535657,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622377204.348, "dur": 233.893,
    "args": {
      "External id": 535658,"Record function id": 0, "Sequence number": 6319509, "Fwd thread id": 1, "Ev Idx": 1129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622377206.256, "dur": 225.414,
    "args": {
      "External id": 535659,"Sequence number": 6319509, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1130
    }
  },
  {
    "ph": "f", "id": 119, "pid": 4183438, "tid": 31367, "ts": 679622377206.256,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622377226.193, "dur": 48.724,
    "args": {
      "External id": 535660,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377229.900, "dur": 3.579,
    "args": {
      "External id": 535661,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622377234.755, "dur": 39.583,
    "args": {
      "External id": 535662,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 1133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622377285.277, "dur": 6.015,
    "args": {
      "External id": 535663,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377287.064, "dur": 3.856,
    "args": {
      "External id": 535664,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622377445.062, "dur": 151.485,
    "args": {
      "External id": 535665,"Record function id": 0, "Sequence number": 6319508, "Fwd thread id": 1, "Ev Idx": 1136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622377447.061, "dur": 142.749,
    "args": {
      "External id": 535666,"Sequence number": 6319508, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1137
    }
  },
  {
    "ph": "f", "id": 120, "pid": 4183438, "tid": 31367, "ts": 679622377447.061,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622377458.629, "dur": 35.473,
    "args": {
      "External id": 535667,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377461.094, "dur": 2.700,
    "args": {
      "External id": 535668,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622377468.627, "dur": 24.894,
    "args": {
      "External id": 535669,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 1140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622377501.378, "dur": 4.144,
    "args": {
      "External id": 535670,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377503.181, "dur": 2.022,
    "args": {
      "External id": 535671,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377602.276, "dur": 14.545,
    "args": {
      "External id": 535672,"Record function id": 0, "Sequence number": 6319507, "Fwd thread id": 1, "Ev Idx": 1143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377604.080, "dur": 10.564,
    "args": {
      "External id": 535673,"Sequence number": 6319507, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1144
    }
  },
  {
    "ph": "f", "id": 121, "pid": 4183438, "tid": 31367, "ts": 679622377604.080,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622377606.433, "dur": 7.944,
    "args": {
      "External id": 535674,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622377607.831, "dur": 6.363,
    "args": {
      "External id": 535675,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377620.336, "dur": 8.034,
    "args": {
      "External id": 535676,"Record function id": 0, "Sequence number": 6319506, "Fwd thread id": 1, "Ev Idx": 1147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377623.166, "dur": 3.667,
    "args": {
      "External id": 535677,"Sequence number": 6319506, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1148
    }
  },
  {
    "ph": "f", "id": 122, "pid": 4183438, "tid": 31367, "ts": 679622377623.166,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622377624.092, "dur": 2.591,
    "args": {
      "External id": 535678,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622377625.640, "dur": 0.887,
    "args": {
      "External id": 535679,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377631.313, "dur": 6.140,
    "args": {
      "External id": 535680,"Record function id": 0, "Sequence number": 6319505, "Fwd thread id": 1, "Ev Idx": 1151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377632.343, "dur": 3.380,
    "args": {
      "External id": 535681,"Sequence number": 6319505, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1152
    }
  },
  {
    "ph": "f", "id": 123, "pid": 4183438, "tid": 31367, "ts": 679622377632.343,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622377633.668, "dur": 1.909,
    "args": {
      "External id": 535682,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622377634.755, "dur": 0.714,
    "args": {
      "External id": 535683,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377640.928, "dur": 7.355,
    "args": {
      "External id": 535684,"Record function id": 0, "Sequence number": 6319504, "Fwd thread id": 1, "Ev Idx": 1155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377642.013, "dur": 4.790,
    "args": {
      "External id": 535685,"Sequence number": 6319504, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1156
    }
  },
  {
    "ph": "f", "id": 124, "pid": 4183438, "tid": 31367, "ts": 679622377642.013,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622377642.958, "dur": 3.706,
    "args": {
      "External id": 535686,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622377645.701, "dur": 0.854,
    "args": {
      "External id": 535687,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377651.542, "dur": 201.657,
    "args": {
      "External id": 535688,"Record function id": 0, "Sequence number": 6319503, "Fwd thread id": 1, "Ev Idx": 1159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377652.542, "dur": 191.836,
    "args": {
      "External id": 535689,"Sequence number": 6319503, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1160
    }
  },
  {
    "ph": "f", "id": 125, "pid": 4183438, "tid": 31367, "ts": 679622377652.542,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622377695.783, "dur": 8.183,
    "args": {
      "External id": 535690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622377697.863, "dur": 5.197,
    "args": {
      "External id": 535691,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377700.751, "dur": 1.837,
    "args": {
      "External id": 535692,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622377705.408, "dur": 83.612,
    "args": {
      "External id": 535693,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622377792.131, "dur": 4.970,
    "args": {
      "External id": 535694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622377793.077, "dur": 3.104,
    "args": {
      "External id": 535695,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377794.861, "dur": 1.166,
    "args": {
      "External id": 535696,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622377798.683, "dur": 3.522,
    "args": {
      "External id": 535697,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622377799.405, "dur": 2.273,
    "args": {
      "External id": 535698,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377800.978, "dur": 0.590,
    "args": {
      "External id": 535699,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622377804.784, "dur": 38.834,
    "args": {
      "External id": 535700,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377862.292, "dur": 8.366,
    "args": {
      "External id": 535701,"Record function id": 0, "Sequence number": 6319502, "Fwd thread id": 1, "Ev Idx": 1172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377863.593, "dur": 5.529,
    "args": {
      "External id": 535702,"Sequence number": 6319502, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1173
    }
  },
  {
    "ph": "f", "id": 126, "pid": 4183438, "tid": 31367, "ts": 679622377863.593,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622377865.632, "dur": 3.350,
    "args": {
      "External id": 535703,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622377866.993, "dur": 1.846,
    "args": {
      "External id": 535704,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377874.039, "dur": 9.853,
    "args": {
      "External id": 535705,"Record function id": 0, "Sequence number": 6319501, "Fwd thread id": 1, "Ev Idx": 1176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377874.910, "dur": 6.769,
    "args": {
      "External id": 535706,"Sequence number": 6319501, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1177
    }
  },
  {
    "ph": "f", "id": 127, "pid": 4183438, "tid": 31367, "ts": 679622377874.910,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622377876.197, "dur": 5.284,
    "args": {
      "External id": 535707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622377876.987, "dur": 4.005,
    "args": {
      "External id": 535708,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377880.284, "dur": 0.630,
    "args": {
      "External id": 535709,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622377889.440, "dur": 9.570,
    "args": {
      "External id": 535710,"Record function id": 0, "Ev Idx": 1181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622377891.469, "dur": 6.850,
    "args": {
      "External id": 535711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622377894.172, "dur": 3.762,
    "args": {
      "External id": 535712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622377895.403, "dur": 2.425,
    "args": {
      "External id": 535713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377902.230, "dur": 6.414,
    "args": {
      "External id": 535714,"Record function id": 0, "Sequence number": 6319500, "Fwd thread id": 1, "Ev Idx": 1185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377903.810, "dur": 2.853,
    "args": {
      "External id": 535715,"Sequence number": 6319500, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1186
    }
  },
  {
    "ph": "f", "id": 128, "pid": 4183438, "tid": 31367, "ts": 679622377903.810,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622377904.808, "dur": 1.712,
    "args": {
      "External id": 535716,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622377905.584, "dur": 0.806,
    "args": {
      "External id": 535717,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377911.912, "dur": 93.623,
    "args": {
      "External id": 535718,"Record function id": 0, "Sequence number": 6319499, "Fwd thread id": 1, "Ev Idx": 1189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622377912.643, "dur": 88.202,
    "args": {
      "External id": 535719,"Sequence number": 6319499, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1190
    }
  },
  {
    "ph": "f", "id": 129, "pid": 4183438, "tid": 31367, "ts": 679622377912.643,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622377916.262, "dur": 3.465,
    "args": {
      "External id": 535720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622377917.028, "dur": 2.184,
    "args": {
      "External id": 535721,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377918.474, "dur": 0.587,
    "args": {
      "External id": 535722,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622377920.367, "dur": 34.698,
    "args": {
      "External id": 535723,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622377955.833, "dur": 4.222,
    "args": {
      "External id": 535724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622377956.299, "dur": 3.022,
    "args": {
      "External id": 535725,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377958.215, "dur": 0.985,
    "args": {
      "External id": 535726,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622377962.741, "dur": 5.334,
    "args": {
      "External id": 535727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622377963.991, "dur": 3.628,
    "args": {
      "External id": 535728,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622377965.515, "dur": 2.015,
    "args": {
      "External id": 535729,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622377968.636, "dur": 31.386,
    "args": {
      "External id": 535730,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378009.468, "dur": 35.372,
    "args": {
      "External id": 535731,"Record function id": 0, "Sequence number": 6319498, "Fwd thread id": 1, "Ev Idx": 1202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378010.727, "dur": 5.738,
    "args": {
      "External id": 535732,"Sequence number": 6319498, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1203
    }
  },
  {
    "ph": "f", "id": 130, "pid": 4183438, "tid": 31367, "ts": 679622378010.727,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378012.057, "dur": 4.260,
    "args": {
      "External id": 535733,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378014.856, "dur": 1.348,
    "args": {
      "External id": 535734,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622378019.376, "dur": 23.303,
    "args": {
      "External id": 535735,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378048.585, "dur": 8.005,
    "args": {
      "External id": 535736,"Record function id": 0, "Sequence number": 6319497, "Fwd thread id": 1, "Ev Idx": 1207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378049.435, "dur": 4.878,
    "args": {
      "External id": 535737,"Sequence number": 6319497, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1208
    }
  },
  {
    "ph": "f", "id": 131, "pid": 4183438, "tid": 31367, "ts": 679622378049.435,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622378050.123, "dur": 4.004,
    "args": {
      "External id": 535738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622378051.202, "dur": 2.429,
    "args": {
      "External id": 535739,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378052.825, "dur": 0.655,
    "args": {
      "External id": 535740,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622378060.473, "dur": 7.872,
    "args": {
      "External id": 535741,"Record function id": 0, "Ev Idx": 1212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622378062.134, "dur": 5.763,
    "args": {
      "External id": 535742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622378063.126, "dur": 4.468,
    "args": {
      "External id": 535743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622378065.984, "dur": 1.512,
    "args": {
      "External id": 535744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378071.582, "dur": 8.477,
    "args": {
      "External id": 535745,"Record function id": 0, "Sequence number": 6319496, "Fwd thread id": 1, "Ev Idx": 1216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378073.049, "dur": 5.114,
    "args": {
      "External id": 535746,"Sequence number": 6319496, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1217
    }
  },
  {
    "ph": "f", "id": 132, "pid": 4183438, "tid": 31367, "ts": 679622378073.049,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378073.903, "dur": 4.100,
    "args": {
      "External id": 535747,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378076.953, "dur": 0.940,
    "args": {
      "External id": 535748,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378083.174, "dur": 122.604,
    "args": {
      "External id": 535749,"Record function id": 0, "Sequence number": 6319495, "Fwd thread id": 1, "Ev Idx": 1220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378083.850, "dur": 113.159,
    "args": {
      "External id": 535750,"Sequence number": 6319495, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1221
    }
  },
  {
    "ph": "f", "id": 133, "pid": 4183438, "tid": 31367, "ts": 679622378083.850,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622378087.892, "dur": 3.266,
    "args": {
      "External id": 535751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622378088.664, "dur": 2.044,
    "args": {
      "External id": 535752,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378090.225, "dur": 0.377,
    "args": {
      "External id": 535753,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622378091.844, "dur": 55.503,
    "args": {
      "External id": 535754,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622378149.510, "dur": 7.893,
    "args": {
      "External id": 535755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622378150.449, "dur": 6.077,
    "args": {
      "External id": 535756,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378153.655, "dur": 2.718,
    "args": {
      "External id": 535757,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622378158.616, "dur": 3.832,
    "args": {
      "External id": 535758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622378159.865, "dur": 2.165,
    "args": {
      "External id": 535759,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378161.287, "dur": 0.673,
    "args": {
      "External id": 535760,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622378163.043, "dur": 33.317,
    "args": {
      "External id": 535761,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378211.527, "dur": 29.594,
    "args": {
      "External id": 535762,"Record function id": 0, "Sequence number": 6319494, "Fwd thread id": 1, "Ev Idx": 1233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378212.569, "dur": 6.091,
    "args": {
      "External id": 535763,"Sequence number": 6319494, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1234
    }
  },
  {
    "ph": "f", "id": 134, "pid": 4183438, "tid": 31367, "ts": 679622378212.569,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378214.283, "dur": 4.237,
    "args": {
      "External id": 535764,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378216.970, "dur": 1.448,
    "args": {
      "External id": 535765,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622378221.147, "dur": 17.529,
    "args": {
      "External id": 535766,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378244.533, "dur": 7.645,
    "args": {
      "External id": 535767,"Record function id": 0, "Sequence number": 6319493, "Fwd thread id": 1, "Ev Idx": 1238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378245.432, "dur": 4.770,
    "args": {
      "External id": 535768,"Sequence number": 6319493, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1239
    }
  },
  {
    "ph": "f", "id": 135, "pid": 4183438, "tid": 31367, "ts": 679622378245.432,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622378246.208, "dur": 3.810,
    "args": {
      "External id": 535769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622378246.958, "dur": 2.531,
    "args": {
      "External id": 535770,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378248.747, "dur": 0.604,
    "args": {
      "External id": 535771,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622378256.124, "dur": 6.011,
    "args": {
      "External id": 535772,"Record function id": 0, "Ev Idx": 1243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622378257.927, "dur": 3.721,
    "args": {
      "External id": 535773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622378259.086, "dur": 2.236,
    "args": {
      "External id": 535774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622378259.648, "dur": 1.579,
    "args": {
      "External id": 535775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622378268.060, "dur": 348.759,
    "args": {
      "External id": 535776,"Record function id": 0, "Sequence number": 6319492, "Fwd thread id": 1, "Ev Idx": 1247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622378269.949, "dur": 317.743,
    "args": {
      "External id": 535777,"Sequence number": 6319492, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1248
    }
  },
  {
    "ph": "f", "id": 136, "pid": 4183438, "tid": 31367, "ts": 679622378269.949,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378303.475, "dur": 2.049,
    "args": {
      "External id": 535778,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378304.126, "dur": 1.282,
    "args": {
      "External id": 535779,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622378319.883, "dur": 5.910,
    "args": {
      "External id": 535780,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622378334.835, "dur": 3.353,
    "args": {
      "External id": 535781,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378484.320, "dur": 2.218,
    "args": {
      "External id": 535782,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 1253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622378490.739, "dur": 35.373,
    "args": {
      "External id": 535783,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 1254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378500.100, "dur": 1.019,
    "args": {
      "External id": 535784,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 1255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622378531.666, "dur": 31.867,
    "args": {
      "External id": 535785,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 1256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622378535.490, "dur": 27.810,
    "args": {
      "External id": 535786,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 1257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378539.686, "dur": 4.280,
    "args": {
      "External id": 535787,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622378545.661, "dur": 16.973,
    "args": {
      "External id": 535788,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 1259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622378567.712, "dur": 3.137,
    "args": {
      "External id": 535789,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 1260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378569.409, "dur": 1.329,
    "args": {
      "External id": 535790,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 1261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378576.906, "dur": 4.213,
    "args": {
      "External id": 535791,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378579.974, "dur": 1.000,
    "args": {
      "External id": 535792,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622378596.506, "dur": 16.646,
    "args": {
      "External id": 535793,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622378625.805, "dur": 8.053,
    "args": {
      "External id": 535794,"Record function id": 0, "Ev Idx": 1265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622378628.536, "dur": 4.661,
    "args": {
      "External id": 535795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622378629.956, "dur": 2.213,
    "args": {
      "External id": 535796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622378630.983, "dur": 1.092,
    "args": {
      "External id": 535797,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378637.422, "dur": 5.272,
    "args": {
      "External id": 535798,"Record function id": 0, "Sequence number": 6319491, "Fwd thread id": 1, "Ev Idx": 1269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622378638.726, "dur": 1.055,
    "args": {
      "External id": 535799,"Sequence number": 6319491, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1270
    }
  },
  {
    "ph": "f", "id": 137, "pid": 4183438, "tid": 31367, "ts": 679622378638.726,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622378646.259, "dur": 435.656,
    "args": {
      "External id": 535800,"Record function id": 0, "Sequence number": 6319490, "Fwd thread id": 1, "Ev Idx": 1271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622378647.405, "dur": 424.779,
    "args": {
      "External id": 535801,"Sequence number": 6319490, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1272
    }
  },
  {
    "ph": "f", "id": 138, "pid": 4183438, "tid": 31367, "ts": 679622378647.405,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378718.608, "dur": 10.313,
    "args": {
      "External id": 535802,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622378724.646, "dur": 3.834,
    "args": {
      "External id": 535803,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]", "[4096, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[16777216, 4096, 1], [], []], "Input Dims": [[16, 4096, 1024], [], []], "Ev Idx": 1274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622378732.038, "dur": 5.411,
    "args": {
      "External id": 535804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622378733.891, "dur": 3.001,
    "args": {
      "External id": 535805,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378735.919, "dur": 0.777,
    "args": {
      "External id": 535806,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 31367,
    "ts": 679622378740.954, "dur": 88.320,
    "args": {
      "External id": 535807,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4096, 1], [1, 2816], []], "Input Dims": [[65536, 1024], [2816, 1024], []], "Ev Idx": 1278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622378743.810, "dur": 5.257,
    "args": {
      "External id": 535808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2816]], "Input Dims": [[2816, 1024]], "Ev Idx": 1279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622378744.288, "dur": 4.285,
    "args": {
      "External id": 535809,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2816], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 1280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378746.002, "dur": 2.490,
    "args": {
      "External id": 535810,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[2816, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2816], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 1281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 31367,
    "ts": 679622378750.465, "dur": 78.222,
    "args": {
      "External id": 535811,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 1282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622378751.863, "dur": 75.944,
    "args": {
      "External id": 535812,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4096, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 1283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622378833.352, "dur": 2.981,
    "args": {
      "External id": 535813,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [11534336, 2816, 1]], "Input Dims": [[65536, 2816], [16, 4096, 2816]], "Ev Idx": 1284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378834.838, "dur": 1.364,
    "args": {
      "External id": 535814,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 1285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622378871.250, "dur": 4.483,
    "args": {
      "External id": 535815,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622378877.283, "dur": 1.701,
    "args": {
      "External id": 535816,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622378880.001, "dur": 1.510,
    "args": {
      "External id": 535817,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378914.304, "dur": 2.531,
    "args": {
      "External id": 535818,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378915.618, "dur": 1.080,
    "args": {
      "External id": 535819,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 4183438, "tid": 31367,
    "ts": 679622378940.457, "dur": 112.990,
    "args": {
      "External id": 535820,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[4096, 1], [2816, 1]], []], "Input Dims": [[], [[65536, 1024], [65536, 2816]], []], "Ev Idx": 1291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622378945.664, "dur": 7.378,
    "args": {
      "External id": 535821,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378949.711, "dur": 2.538,
    "args": {
      "External id": 535822,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024, 1]", "[4096, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622378955.067, "dur": 6.645,
    "args": {
      "External id": 535823,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4096, 1, 1], []], "Input Dims": [[65536, 1024, 1], []], "Ev Idx": 1294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378958.522, "dur": 2.372,
    "args": {
      "External id": 535824,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 65536]", "[1, 1, 4096]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1, 1], [], [], []], "Input Dims": [[65536, 1024, 1], [], [], []], "Ev Idx": 1295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 31367,
    "ts": 679622378962.950, "dur": 2.797,
    "args": {
      "External id": 535825,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 1296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378964.827, "dur": 0.506,
    "args": {
      "External id": 535826,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 1297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622378966.890, "dur": 3.660,
    "args": {
      "External id": 535827,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 1298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378969.093, "dur": 0.783,
    "args": {
      "External id": 535828,"Record function id": 0, "Concrete Inputs": ["", "[1, 2816, 65536]", "[1, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1, 1], [], [], []], "Input Dims": [[65536, 2816, 1], [], [], []], "Ev Idx": 1299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622378974.559, "dur": 3.421,
    "args": {
      "External id": 535829,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 4096], []], "Input Dims": [[1024, 1, 65536], []], "Ev Idx": 1300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378976.951, "dur": 0.692,
    "args": {
      "External id": 535830,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536, 1]", "[1, 4096, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 4096], [], [], []], "Input Dims": [[1024, 1, 65536], [], [], []], "Ev Idx": 1301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378979.134, "dur": 5.692,
    "args": {
      "External id": 535831,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 4096, 1], []], "Input Dims": [[1024, 65536, 1], []], "Ev Idx": 1302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 4183438, "tid": 31367,
    "ts": 679622378982.953, "dur": 1.683,
    "args": {
      "External id": 535832,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024, 65536]", "[1024, 1, 4096]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 4096, 1], [], []], "Input Dims": [[1024, 65536, 1], [], []], "Ev Idx": 1303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622378987.776, "dur": 2.578,
    "args": {
      "External id": 535833,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 2816], []], "Input Dims": [[1, 2816, 65536], []], "Ev Idx": 1304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622378989.587, "dur": 0.448,
    "args": {
      "External id": 535834,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816, 1]", "[2816, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 2816], [], [], []], "Input Dims": [[1, 2816, 65536], [], [], []], "Ev Idx": 1305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622378991.182, "dur": 3.378,
    "args": {
      "External id": 535835,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 1306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622378992.288, "dur": 2.192,
    "args": {
      "External id": 535836,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 1], []], "Input Dims": [[65536, 2816, 1], []], "Ev Idx": 1307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622378995.810, "dur": 44.580,
    "args": {
      "External id": 535837,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1, 4096], [184549376, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816]], "Ev Idx": 1308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379042.123, "dur": 2.949,
    "args": {
      "External id": 535838,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2883584, 2816, 1], []], "Input Dims": [[1, 1024, 2816], []], "Ev Idx": 1309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 31367,
    "ts": 679622379046.096, "dur": 3.150,
    "args": {
      "External id": 535839,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 2816, 1], []], "Input Dims": [[1024, 1, 2816], []], "Ev Idx": 1310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379048.236, "dur": 0.506,
    "args": {
      "External id": 535840,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816, 1]", "[2816, 1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 2816, 1], [], [], []], "Input Dims": [[1024, 1, 2816], [], [], []], "Ev Idx": 1311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379051.501, "dur": 0.968,
    "args": {
      "External id": 535841,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1, 2816], []], "Input Dims": [[1024, 2816, 1], []], "Ev Idx": 1312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379093.283, "dur": 28.429,
    "args": {
      "External id": 535842,"Record function id": 0, "Ev Idx": 1313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379095.552, "dur": 24.928,
    "args": {
      "External id": 535843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379114.407, "dur": 4.665,
    "args": {
      "External id": 535844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379116.136, "dur": 2.672,
    "args": {
      "External id": 535845,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379127.285, "dur": 7.532,
    "args": {
      "External id": 535846,"Record function id": 0, "Sequence number": 6319489, "Fwd thread id": 1, "Ev Idx": 1317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379128.781, "dur": 4.156,
    "args": {
      "External id": 535847,"Sequence number": 6319489, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 1318
    }
  },
  {
    "ph": "f", "id": 139, "pid": 4183438, "tid": 31367, "ts": 679622379128.781,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379130.655, "dur": 2.066,
    "args": {
      "External id": 535848,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379131.540, "dur": 1.053,
    "args": {
      "External id": 535849,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379138.253, "dur": 109.998,
    "args": {
      "External id": 535850,"Record function id": 0, "Sequence number": 6319488, "Fwd thread id": 1, "Ev Idx": 1321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379139.168, "dur": 103.280,
    "args": {
      "External id": 535851,"Sequence number": 6319488, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 1322
    }
  },
  {
    "ph": "f", "id": 140, "pid": 4183438, "tid": 31367, "ts": 679622379139.168,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379144.044, "dur": 4.833,
    "args": {
      "External id": 535852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 1323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379145.311, "dur": 2.944,
    "args": {
      "External id": 535853,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 1324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379147.121, "dur": 0.943,
    "args": {
      "External id": 535854,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 1325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622379149.766, "dur": 39.423,
    "args": {
      "External id": 535855,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 1326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379190.311, "dur": 6.040,
    "args": {
      "External id": 535856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379191.221, "dur": 4.479,
    "args": {
      "External id": 535857,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 1328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379194.517, "dur": 1.014,
    "args": {
      "External id": 535858,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 1329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379197.604, "dur": 3.483,
    "args": {
      "External id": 535859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379198.448, "dur": 2.151,
    "args": {
      "External id": 535860,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379200.001, "dur": 0.520,
    "args": {
      "External id": 535861,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622379201.458, "dur": 40.320,
    "args": {
      "External id": 535862,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 1333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379252.748, "dur": 8.632,
    "args": {
      "External id": 535863,"Record function id": 0, "Sequence number": 6319487, "Fwd thread id": 1, "Ev Idx": 1334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379253.773, "dur": 6.136,
    "args": {
      "External id": 535864,"Sequence number": 6319487, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1335
    }
  },
  {
    "ph": "f", "id": 141, "pid": 4183438, "tid": 31367, "ts": 679622379253.773,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379255.410, "dur": 4.359,
    "args": {
      "External id": 535865,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379258.225, "dur": 1.411,
    "args": {
      "External id": 535866,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379264.736, "dur": 7.749,
    "args": {
      "External id": 535867,"Record function id": 0, "Sequence number": 6319486, "Fwd thread id": 1, "Ev Idx": 1338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379265.637, "dur": 5.126,
    "args": {
      "External id": 535868,"Sequence number": 6319486, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1339
    }
  },
  {
    "ph": "f", "id": 142, "pid": 4183438, "tid": 31367, "ts": 679622379265.637,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379266.630, "dur": 3.919,
    "args": {
      "External id": 535869,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379267.491, "dur": 2.544,
    "args": {
      "External id": 535870,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379269.421, "dur": 0.495,
    "args": {
      "External id": 535871,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379276.403, "dur": 5.912,
    "args": {
      "External id": 535872,"Record function id": 0, "Ev Idx": 1343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379278.348, "dur": 3.544,
    "args": {
      "External id": 535873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379279.370, "dur": 2.265,
    "args": {
      "External id": 535874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379280.080, "dur": 1.444,
    "args": {
      "External id": 535875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379285.651, "dur": 7.731,
    "args": {
      "External id": 535876,"Record function id": 0, "Sequence number": 6319485, "Fwd thread id": 1, "Ev Idx": 1347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379286.703, "dur": 4.859,
    "args": {
      "External id": 535877,"Sequence number": 6319485, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[11534336, 2816, 1]], "Input Dims": [[16, 4096, 2816]], "Ev Idx": 1348
    }
  },
  {
    "ph": "f", "id": 143, "pid": 4183438, "tid": 31367, "ts": 679622379286.703,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379290.156, "dur": 1.267,
    "args": {
      "External id": 535878,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379290.689, "dur": 0.607,
    "args": {
      "External id": 535879,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 1350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379296.332, "dur": 87.428,
    "args": {
      "External id": 535880,"Record function id": 0, "Sequence number": 6319484, "Fwd thread id": 1, "Ev Idx": 1351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379297.246, "dur": 79.664,
    "args": {
      "External id": 535881,"Sequence number": 6319484, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 1352
    }
  },
  {
    "ph": "f", "id": 144, "pid": 4183438, "tid": 31367, "ts": 679622379297.246,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379298.995, "dur": 4.324,
    "args": {
      "External id": 535882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[65536, 2816]], "Ev Idx": 1353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379299.407, "dur": 3.462,
    "args": {
      "External id": 535883,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[65536, 2816], [], []], "Ev Idx": 1354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379302.157, "dur": 0.568,
    "args": {
      "External id": 535884,"Record function id": 0, "Concrete Inputs": ["", "[2816, 65536]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[65536, 2816], [], [], []], "Ev Idx": 1355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622379303.958, "dur": 27.850,
    "args": {
      "External id": 535885,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024]], "Ev Idx": 1356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379332.908, "dur": 3.829,
    "args": {
      "External id": 535886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379333.526, "dur": 2.711,
    "args": {
      "External id": 535887,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 1358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379335.215, "dur": 0.873,
    "args": {
      "External id": 535888,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 1359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379337.764, "dur": 6.400,
    "args": {
      "External id": 535889,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379338.730, "dur": 4.991,
    "args": {
      "External id": 535890,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379341.576, "dur": 2.058,
    "args": {
      "External id": 535891,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622379344.629, "dur": 31.677,
    "args": {
      "External id": 535892,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 1363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379389.662, "dur": 31.798,
    "args": {
      "External id": 535893,"Record function id": 0, "Sequence number": 6319483, "Fwd thread id": 1, "Ev Idx": 1364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379390.659, "dur": 4.522,
    "args": {
      "External id": 535894,"Sequence number": 6319483, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1365
    }
  },
  {
    "ph": "f", "id": 145, "pid": 4183438, "tid": 31367, "ts": 679622379390.659,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379392.179, "dur": 2.856,
    "args": {
      "External id": 535895,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379393.436, "dur": 1.444,
    "args": {
      "External id": 535896,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622379398.026, "dur": 20.541,
    "args": {
      "External id": 535897,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379425.160, "dur": 8.739,
    "args": {
      "External id": 535898,"Record function id": 0, "Sequence number": 6319482, "Fwd thread id": 1, "Ev Idx": 1369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379426.046, "dur": 6.081,
    "args": {
      "External id": 535899,"Sequence number": 6319482, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1370
    }
  },
  {
    "ph": "f", "id": 146, "pid": 4183438, "tid": 31367, "ts": 679622379426.046,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379426.918, "dur": 4.979,
    "args": {
      "External id": 535900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 2816]], "Ev Idx": 1371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379427.843, "dur": 3.489,
    "args": {
      "External id": 535901,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 1372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379430.566, "dur": 0.646,
    "args": {
      "External id": 535902,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 1373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379437.903, "dur": 4.621,
    "args": {
      "External id": 535903,"Record function id": 0, "Ev Idx": 1374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379439.685, "dur": 2.432,
    "args": {
      "External id": 535904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379440.524, "dur": 1.362,
    "args": {
      "External id": 535905,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379440.905, "dur": 0.865,
    "args": {
      "External id": 535906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622379446.296, "dur": 476.225,
    "args": {
      "External id": 535907,"Record function id": 0, "Sequence number": 6319481, "Fwd thread id": 1, "Ev Idx": 1378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622379452.311, "dur": 433.502,
    "args": {
      "External id": 535908,"Sequence number": 6319481, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 1379
    }
  },
  {
    "ph": "f", "id": 147, "pid": 4183438, "tid": 31367, "ts": 679622379452.311,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 31367,
    "ts": 679622379475.967, "dur": 35.423,
    "args": {
      "External id": 535909,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622379477.402, "dur": 33.798,
    "args": {
      "External id": 535910,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622379480.438, "dur": 6.191,
    "args": {
      "External id": 535911,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], [], []], "Ev Idx": 1382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622379483.175, "dur": 2.982,
    "args": {
      "External id": 535912,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622379487.927, "dur": 22.703,
    "args": {
      "External id": 535913,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379523.570, "dur": 2.284,
    "args": {
      "External id": 535914,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379524.419, "dur": 1.340,
    "args": {
      "External id": 535915,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379529.852, "dur": 4.012,
    "args": {
      "External id": 535916,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379530.594, "dur": 3.136,
    "args": {
      "External id": 535917,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622379547.341, "dur": 2.708,
    "args": {
      "External id": 535918,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622379561.394, "dur": 3.802,
    "args": {
      "External id": 535919,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379767.741, "dur": 4.991,
    "args": {
      "External id": 535920,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 1391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622379777.470, "dur": 36.059,
    "args": {
      "External id": 535921,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 1392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379788.353, "dur": 0.846,
    "args": {
      "External id": 535922,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 1393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622379819.814, "dur": 31.661,
    "args": {
      "External id": 535923,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 1394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622379821.566, "dur": 29.673,
    "args": {
      "External id": 535924,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 1395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379826.539, "dur": 4.394,
    "args": {
      "External id": 535925,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622379834.140, "dur": 16.453,
    "args": {
      "External id": 535926,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 1397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622379856.318, "dur": 3.124,
    "args": {
      "External id": 535927,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 1398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379857.810, "dur": 1.516,
    "args": {
      "External id": 535928,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 1399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379866.889, "dur": 2.381,
    "args": {
      "External id": 535929,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379867.980, "dur": 1.182,
    "args": {
      "External id": 535930,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379871.644, "dur": 3.684,
    "args": {
      "External id": 535931,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379872.612, "dur": 2.625,
    "args": {
      "External id": 535932,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622379902.813, "dur": 18.217,
    "args": {
      "External id": 535933,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379937.308, "dur": 9.151,
    "args": {
      "External id": 535934,"Record function id": 0, "Ev Idx": 1405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622379939.524, "dur": 6.191,
    "args": {
      "External id": 535935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379941.673, "dur": 3.077,
    "args": {
      "External id": 535936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622379942.662, "dur": 1.939,
    "args": {
      "External id": 535937,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379950.334, "dur": 8.202,
    "args": {
      "External id": 535938,"Record function id": 0, "Sequence number": 6319480, "Fwd thread id": 1, "Ev Idx": 1409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379951.725, "dur": 4.443,
    "args": {
      "External id": 535939,"Sequence number": 6319480, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1410
    }
  },
  {
    "ph": "f", "id": 148, "pid": 4183438, "tid": 31367, "ts": 679622379951.725,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622379953.375, "dur": 2.604,
    "args": {
      "External id": 535940,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622379954.588, "dur": 1.242,
    "args": {
      "External id": 535941,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379963.870, "dur": 160.712,
    "args": {
      "External id": 535942,"Record function id": 0, "Sequence number": 6319479, "Fwd thread id": 1, "Ev Idx": 1413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622379964.717, "dur": 151.125,
    "args": {
      "External id": 535943,"Sequence number": 6319479, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1414
    }
  },
  {
    "ph": "f", "id": 149, "pid": 4183438, "tid": 31367, "ts": 679622379964.717,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622379968.201, "dur": 4.974,
    "args": {
      "External id": 535944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622379969.538, "dur": 3.068,
    "args": {
      "External id": 535945,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622379971.424, "dur": 0.917,
    "args": {
      "External id": 535946,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622379974.179, "dur": 73.678,
    "args": {
      "External id": 535947,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622380049.210, "dur": 7.559,
    "args": {
      "External id": 535948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622380051.934, "dur": 4.208,
    "args": {
      "External id": 535949,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380053.605, "dur": 2.325,
    "args": {
      "External id": 535950,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622380058.398, "dur": 3.831,
    "args": {
      "External id": 535951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622380059.421, "dur": 2.149,
    "args": {
      "External id": 535952,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380061.158, "dur": 0.325,
    "args": {
      "External id": 535953,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622380062.740, "dur": 34.270,
    "args": {
      "External id": 535954,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380131.895, "dur": 11.660,
    "args": {
      "External id": 535955,"Record function id": 0, "Sequence number": 6319478, "Fwd thread id": 1, "Ev Idx": 1426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380133.042, "dur": 8.819,
    "args": {
      "External id": 535956,"Sequence number": 6319478, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1427
    }
  },
  {
    "ph": "f", "id": 150, "pid": 4183438, "tid": 31367, "ts": 679622380133.042,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622380136.825, "dur": 4.873,
    "args": {
      "External id": 535957,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622380138.379, "dur": 3.170,
    "args": {
      "External id": 535958,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380147.285, "dur": 7.150,
    "args": {
      "External id": 535959,"Record function id": 0, "Sequence number": 6319477, "Fwd thread id": 1, "Ev Idx": 1430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380148.141, "dur": 4.560,
    "args": {
      "External id": 535960,"Sequence number": 6319477, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1431
    }
  },
  {
    "ph": "f", "id": 151, "pid": 4183438, "tid": 31367, "ts": 679622380148.141,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622380149.005, "dur": 3.493,
    "args": {
      "External id": 535961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622380149.649, "dur": 2.327,
    "args": {
      "External id": 535962,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380151.269, "dur": 0.595,
    "args": {
      "External id": 535963,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622380158.588, "dur": 8.775,
    "args": {
      "External id": 535964,"Record function id": 0, "Ev Idx": 1435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622380160.489, "dur": 6.347,
    "args": {
      "External id": 535965,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622380161.881, "dur": 4.644,
    "args": {
      "External id": 535966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622380164.232, "dur": 2.156,
    "args": {
      "External id": 535967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380170.469, "dur": 5.556,
    "args": {
      "External id": 535968,"Record function id": 0, "Sequence number": 6319476, "Fwd thread id": 1, "Ev Idx": 1439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380171.212, "dur": 3.376,
    "args": {
      "External id": 535969,"Sequence number": 6319476, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1440
    }
  },
  {
    "ph": "f", "id": 152, "pid": 4183438, "tid": 31367, "ts": 679622380171.212,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622380172.395, "dur": 2.036,
    "args": {
      "External id": 535970,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622380173.036, "dur": 1.257,
    "args": {
      "External id": 535971,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622380180.001, "dur": 276.555,
    "args": {
      "External id": 535972,"Record function id": 0, "Sequence number": 6319475, "Fwd thread id": 1, "Ev Idx": 1443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622380181.727, "dur": 258.994,
    "args": {
      "External id": 535973,"Sequence number": 6319475, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1444
    }
  },
  {
    "ph": "f", "id": 153, "pid": 4183438, "tid": 31367, "ts": 679622380181.727,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622380197.527, "dur": 7.730,
    "args": {
      "External id": 535974,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380200.589, "dur": 4.211,
    "args": {
      "External id": 535975,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622380207.329, "dur": 3.247,
    "args": {
      "External id": 535976,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380208.694, "dur": 1.709,
    "args": {
      "External id": 535977,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622380211.970, "dur": 3.621,
    "args": {
      "External id": 535978,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380214.085, "dur": 1.341,
    "args": {
      "External id": 535979,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622380240.291, "dur": 175.794,
    "args": {
      "External id": 535980,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622380321.054, "dur": 3.999,
    "args": {
      "External id": 535981,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622380326.790, "dur": 3.740,
    "args": {
      "External id": 535982,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622380427.955, "dur": 3.624,
    "args": {
      "External id": 535983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622380434.567, "dur": 0.847,
    "args": {
      "External id": 535984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 31367,
    "ts": 679622380437.434, "dur": 0.589,
    "args": {
      "External id": 535985,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622380463.662, "dur": 270.015,
    "args": {
      "External id": 535986,"Record function id": 0, "Sequence number": 6319474, "Fwd thread id": 1, "Ev Idx": 1457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622380465.521, "dur": 259.494,
    "args": {
      "External id": 535987,"Sequence number": 6319474, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1458
    }
  },
  {
    "ph": "f", "id": 154, "pid": 4183438, "tid": 31367, "ts": 679622380465.521,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622380484.642, "dur": 49.926,
    "args": {
      "External id": 535988,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380487.595, "dur": 2.892,
    "args": {
      "External id": 535989,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622380491.827, "dur": 42.182,
    "args": {
      "External id": 535990,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 1461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622380544.285, "dur": 4.483,
    "args": {
      "External id": 535991,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380546.085, "dur": 2.373,
    "args": {
      "External id": 535992,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622380743.830, "dur": 159.696,
    "args": {
      "External id": 535993,"Record function id": 0, "Sequence number": 6319473, "Fwd thread id": 1, "Ev Idx": 1464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622380746.203, "dur": 151.331,
    "args": {
      "External id": 535994,"Sequence number": 6319473, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1465
    }
  },
  {
    "ph": "f", "id": 155, "pid": 4183438, "tid": 31367, "ts": 679622380746.203,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 31367,
    "ts": 679622380760.393, "dur": 37.136,
    "args": {
      "External id": 535995,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380763.241, "dur": 3.672,
    "args": {
      "External id": 535996,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622380768.044, "dur": 28.819,
    "args": {
      "External id": 535997,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], []], "Ev Idx": 1468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 31367,
    "ts": 679622380805.547, "dur": 5.144,
    "args": {
      "External id": 535998,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 1469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380807.507, "dur": 2.920,
    "args": {
      "External id": 535999,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380909.681, "dur": 16.577,
    "args": {
      "External id": 536000,"Record function id": 0, "Sequence number": 6319472, "Fwd thread id": 1, "Ev Idx": 1471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380910.846, "dur": 13.339,
    "args": {
      "External id": 536001,"Sequence number": 6319472, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1472
    }
  },
  {
    "ph": "f", "id": 156, "pid": 4183438, "tid": 31367, "ts": 679622380910.846,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622380915.671, "dur": 8.205,
    "args": {
      "External id": 536002,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622380917.507, "dur": 6.215,
    "args": {
      "External id": 536003,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380929.834, "dur": 6.494,
    "args": {
      "External id": 536004,"Record function id": 0, "Sequence number": 6319471, "Fwd thread id": 1, "Ev Idx": 1475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380930.651, "dur": 3.875,
    "args": {
      "External id": 536005,"Sequence number": 6319471, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1476
    }
  },
  {
    "ph": "f", "id": 157, "pid": 4183438, "tid": 31367, "ts": 679622380930.651,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622380932.157, "dur": 2.202,
    "args": {
      "External id": 536006,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622380933.247, "dur": 0.968,
    "args": {
      "External id": 536007,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380939.221, "dur": 5.708,
    "args": {
      "External id": 536008,"Record function id": 0, "Sequence number": 6319470, "Fwd thread id": 1, "Ev Idx": 1479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380939.941, "dur": 3.326,
    "args": {
      "External id": 536009,"Sequence number": 6319470, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 1480
    }
  },
  {
    "ph": "f", "id": 158, "pid": 4183438, "tid": 31367, "ts": 679622380939.941,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622380941.203, "dur": 1.905,
    "args": {
      "External id": 536010,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622380942.149, "dur": 0.813,
    "args": {
      "External id": 536011,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 1482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380948.473, "dur": 8.187,
    "args": {
      "External id": 536012,"Record function id": 0, "Sequence number": 6319469, "Fwd thread id": 1, "Ev Idx": 1483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380949.779, "dur": 4.694,
    "args": {
      "External id": 536013,"Sequence number": 6319469, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1484
    }
  },
  {
    "ph": "f", "id": 159, "pid": 4183438, "tid": 31367, "ts": 679622380949.779,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622380952.305, "dur": 2.013,
    "args": {
      "External id": 536014,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622380953.462, "dur": 0.741,
    "args": {
      "External id": 536015,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380959.767, "dur": 178.013,
    "args": {
      "External id": 536016,"Record function id": 0, "Sequence number": 6319468, "Fwd thread id": 1, "Ev Idx": 1487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622380960.607, "dur": 168.621,
    "args": {
      "External id": 536017,"Sequence number": 6319468, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1488
    }
  },
  {
    "ph": "f", "id": 160, "pid": 4183438, "tid": 31367, "ts": 679622380960.607,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622380963.815, "dur": 8.673,
    "args": {
      "External id": 536018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622380965.887, "dur": 5.958,
    "args": {
      "External id": 536019,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622380969.849, "dur": 1.760,
    "args": {
      "External id": 536020,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622380973.867, "dur": 81.704,
    "args": {
      "External id": 536021,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381056.993, "dur": 5.304,
    "args": {
      "External id": 536022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381058.108, "dur": 3.533,
    "args": {
      "External id": 536023,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381060.219, "dur": 1.241,
    "args": {
      "External id": 536024,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381063.884, "dur": 5.395,
    "args": {
      "External id": 536025,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381065.085, "dur": 3.690,
    "args": {
      "External id": 536026,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381068.379, "dur": 0.325,
    "args": {
      "External id": 536027,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622381069.932, "dur": 57.819,
    "args": {
      "External id": 536028,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381145.448, "dur": 8.316,
    "args": {
      "External id": 536029,"Record function id": 0, "Sequence number": 6319467, "Fwd thread id": 1, "Ev Idx": 1500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381146.655, "dur": 5.422,
    "args": {
      "External id": 536030,"Sequence number": 6319467, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1501
    }
  },
  {
    "ph": "f", "id": 161, "pid": 4183438, "tid": 31367, "ts": 679622381146.655,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622381148.599, "dur": 3.321,
    "args": {
      "External id": 536031,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381149.946, "dur": 1.865,
    "args": {
      "External id": 536032,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381157.007, "dur": 10.104,
    "args": {
      "External id": 536033,"Record function id": 0, "Sequence number": 6319466, "Fwd thread id": 1, "Ev Idx": 1504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381157.698, "dur": 7.239,
    "args": {
      "External id": 536034,"Sequence number": 6319466, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1505
    }
  },
  {
    "ph": "f", "id": 162, "pid": 4183438, "tid": 31367, "ts": 679622381157.698,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381158.509, "dur": 6.207,
    "args": {
      "External id": 536035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381161.762, "dur": 2.382,
    "args": {
      "External id": 536036,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381163.515, "dur": 0.511,
    "args": {
      "External id": 536037,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381172.737, "dur": 9.284,
    "args": {
      "External id": 536038,"Record function id": 0, "Ev Idx": 1509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381174.543, "dur": 6.811,
    "args": {
      "External id": 536039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381177.115, "dur": 3.802,
    "args": {
      "External id": 536040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381178.305, "dur": 2.509,
    "args": {
      "External id": 536041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381185.429, "dur": 6.232,
    "args": {
      "External id": 536042,"Record function id": 0, "Sequence number": 6319465, "Fwd thread id": 1, "Ev Idx": 1513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381187.113, "dur": 2.851,
    "args": {
      "External id": 536043,"Sequence number": 6319465, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1514
    }
  },
  {
    "ph": "f", "id": 163, "pid": 4183438, "tid": 31367, "ts": 679622381187.113,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622381188.102, "dur": 1.705,
    "args": {
      "External id": 536044,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381188.737, "dur": 0.980,
    "args": {
      "External id": 536045,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381196.520, "dur": 99.614,
    "args": {
      "External id": 536046,"Record function id": 0, "Sequence number": 6319464, "Fwd thread id": 1, "Ev Idx": 1517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381197.522, "dur": 93.249,
    "args": {
      "External id": 536047,"Sequence number": 6319464, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1518
    }
  },
  {
    "ph": "f", "id": 164, "pid": 4183438, "tid": 31367, "ts": 679622381197.522,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381199.330, "dur": 3.121,
    "args": {
      "External id": 536048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381199.810, "dur": 2.195,
    "args": {
      "External id": 536049,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381201.107, "dur": 0.783,
    "args": {
      "External id": 536050,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622381203.281, "dur": 39.634,
    "args": {
      "External id": 536051,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381244.009, "dur": 5.741,
    "args": {
      "External id": 536052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381244.679, "dur": 4.562,
    "args": {
      "External id": 536053,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381247.970, "dur": 1.151,
    "args": {
      "External id": 536054,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381251.125, "dur": 5.974,
    "args": {
      "External id": 536055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381252.557, "dur": 4.146,
    "args": {
      "External id": 536056,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381254.281, "dur": 2.353,
    "args": {
      "External id": 536057,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622381257.760, "dur": 32.352,
    "args": {
      "External id": 536058,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381302.150, "dur": 37.864,
    "args": {
      "External id": 536059,"Record function id": 0, "Sequence number": 6319463, "Fwd thread id": 1, "Ev Idx": 1530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381303.186, "dur": 8.662,
    "args": {
      "External id": 536060,"Sequence number": 6319463, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1531
    }
  },
  {
    "ph": "f", "id": 165, "pid": 4183438, "tid": 31367, "ts": 679622381303.186,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622381306.815, "dur": 4.886,
    "args": {
      "External id": 536061,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381310.152, "dur": 1.436,
    "args": {
      "External id": 536062,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 31367,
    "ts": 679622381315.093, "dur": 22.556,
    "args": {
      "External id": 536063,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381343.583, "dur": 27.390,
    "args": {
      "External id": 536064,"Record function id": 0, "Sequence number": 6319462, "Fwd thread id": 1, "Ev Idx": 1535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381363.298, "dur": 5.946,
    "args": {
      "External id": 536065,"Sequence number": 6319462, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1536
    }
  },
  {
    "ph": "f", "id": 166, "pid": 4183438, "tid": 31367, "ts": 679622381363.298,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381364.039, "dur": 4.999,
    "args": {
      "External id": 536066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381364.794, "dur": 3.719,
    "args": {
      "External id": 536067,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381367.993, "dur": 0.408,
    "args": {
      "External id": 536068,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381374.830, "dur": 5.034,
    "args": {
      "External id": 536069,"Record function id": 0, "Ev Idx": 1540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381376.238, "dur": 3.103,
    "args": {
      "External id": 536070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381377.201, "dur": 1.866,
    "args": {
      "External id": 536071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381377.688, "dur": 1.227,
    "args": {
      "External id": 536072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381383.032, "dur": 5.311,
    "args": {
      "External id": 536073,"Record function id": 0, "Sequence number": 6319461, "Fwd thread id": 1, "Ev Idx": 1544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381384.212, "dur": 2.630,
    "args": {
      "External id": 536074,"Sequence number": 6319461, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1545
    }
  },
  {
    "ph": "f", "id": 167, "pid": 4183438, "tid": 31367, "ts": 679622381384.212,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622381385.150, "dur": 1.553,
    "args": {
      "External id": 536075,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381385.799, "dur": 0.806,
    "args": {
      "External id": 536076,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381391.456, "dur": 91.840,
    "args": {
      "External id": 536077,"Record function id": 0, "Sequence number": 6319460, "Fwd thread id": 1, "Ev Idx": 1548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381392.071, "dur": 84.383,
    "args": {
      "External id": 536078,"Sequence number": 6319460, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1549
    }
  },
  {
    "ph": "f", "id": 168, "pid": 4183438, "tid": 31367, "ts": 679622381392.071,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381395.718, "dur": 2.895,
    "args": {
      "External id": 536079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381396.170, "dur": 1.994,
    "args": {
      "External id": 536080,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[65536, 1024], [], []], "Ev Idx": 1551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381397.624, "dur": 0.442,
    "args": {
      "External id": 536081,"Record function id": 0, "Concrete Inputs": ["", "[1024, 65536]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 1552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622381399.329, "dur": 35.501,
    "args": {
      "External id": 536082,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024]], "Ev Idx": 1553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381435.760, "dur": 5.367,
    "args": {
      "External id": 536083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381436.557, "dur": 3.915,
    "args": {
      "External id": 536084,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381439.933, "dur": 0.420,
    "args": {
      "External id": 536085,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381442.045, "dur": 3.098,
    "args": {
      "External id": 536086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381442.856, "dur": 1.776,
    "args": {
      "External id": 536087,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381444.114, "dur": 0.443,
    "args": {
      "External id": 536088,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622381445.578, "dur": 30.258,
    "args": {
      "External id": 536089,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 1560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381487.556, "dur": 23.527,
    "args": {
      "External id": 536090,"Record function id": 0, "Sequence number": 6319459, "Fwd thread id": 1, "Ev Idx": 1561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381488.306, "dur": 4.168,
    "args": {
      "External id": 536091,"Sequence number": 6319459, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 1562
    }
  },
  {
    "ph": "f", "id": 169, "pid": 4183438, "tid": 31367, "ts": 679622381488.306,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622381490.163, "dur": 2.154,
    "args": {
      "External id": 536092,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381491.100, "dur": 1.067,
    "args": {
      "External id": 536093,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622381494.641, "dur": 14.382,
    "args": {
      "External id": 536094,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381514.541, "dur": 8.813,
    "args": {
      "External id": 536095,"Record function id": 0, "Sequence number": 6319458, "Fwd thread id": 1, "Ev Idx": 1566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 4183438, "tid": 31367,
    "ts": 679622381517.139, "dur": 3.935,
    "args": {
      "External id": 536096,"Sequence number": 6319458, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1567
    }
  },
  {
    "ph": "f", "id": 170, "pid": 4183438, "tid": 31367, "ts": 679622381517.139,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 31367,
    "ts": 679622381517.741, "dur": 3.128,
    "args": {
      "External id": 536097,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 1024]], "Input Dims": [[1024, 1024]], "Ev Idx": 1568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 31367,
    "ts": 679622381518.340, "dur": 1.993,
    "args": {
      "External id": 536098,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 1024], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 1569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381519.794, "dur": 0.418,
    "args": {
      "External id": 536099,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1024, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1024], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 1570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381527.028, "dur": 5.183,
    "args": {
      "External id": 536100,"Record function id": 0, "Ev Idx": 1571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381528.881, "dur": 2.909,
    "args": {
      "External id": 536101,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381529.708, "dur": 1.806,
    "args": {
      "External id": 536102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381530.381, "dur": 1.027,
    "args": {
      "External id": 536103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622381536.102, "dur": 389.959,
    "args": {
      "External id": 536104,"Record function id": 0, "Sequence number": 6319457, "Fwd thread id": 1, "Ev Idx": 1575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622381537.497, "dur": 354.686,
    "args": {
      "External id": 536105,"Sequence number": 6319457, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1576
    }
  },
  {
    "ph": "f", "id": 171, "pid": 4183438, "tid": 31367, "ts": 679622381537.497,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622381568.300, "dur": 3.221,
    "args": {
      "External id": 536106,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381570.368, "dur": 0.979,
    "args": {
      "External id": 536107,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 1578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622381586.635, "dur": 6.032,
    "args": {
      "External id": 536108,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622381601.999, "dur": 3.150,
    "args": {
      "External id": 536109,"Record function id": 0, "Concrete Inputs": ["[132, 1024]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381791.412, "dur": 3.038,
    "args": {
      "External id": 536110,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 1024]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[132, 1024], []], "Ev Idx": 1581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 31367,
    "ts": 679622381798.868, "dur": 36.321,
    "args": {
      "External id": 536111,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[135168, 1024, 1], [], [], []], "Input Dims": [[1, 132, 1024], [], [], []], "Ev Idx": 1582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381810.521, "dur": 1.079,
    "args": {
      "External id": 536112,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 1024]", "[1024, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1, 1024], [], [], []], "Ev Idx": 1583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622381840.820, "dur": 29.571,
    "args": {
      "External id": 536113,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], []], "Ev Idx": 1584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622381844.159, "dur": 26.000,
    "args": {
      "External id": 536114,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[1, 1024], [], [], [], [], [], []], "Ev Idx": 1585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622381848.634, "dur": 3.899,
    "args": {
      "External id": 536115,"Record function id": 0, "Concrete Inputs": ["[1, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622381854.101, "dur": 15.507,
    "args": {
      "External id": 536116,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[1, 1024], [1, 1024], []], "Ev Idx": 1587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 31367,
    "ts": 679622381874.746, "dur": 2.669,
    "args": {
      "External id": 536117,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1]], "Input Dims": [[1, 1024], [1024]], "Ev Idx": 1588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381876.014, "dur": 1.285,
    "args": {
      "External id": 536118,"Record function id": 0, "Concrete Inputs": ["", "[1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[1, 1024], []], "Ev Idx": 1589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 31367,
    "ts": 679622381883.517, "dur": 2.057,
    "args": {
      "External id": 536119,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622381884.768, "dur": 0.707,
    "args": {
      "External id": 536120,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 1591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 31367,
    "ts": 679622381906.452, "dur": 15.034,
    "args": {
      "External id": 536121,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 1592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381938.677, "dur": 8.772,
    "args": {
      "External id": 536122,"Record function id": 0, "Ev Idx": 1593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622381941.059, "dur": 5.672,
    "args": {
      "External id": 536123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381942.924, "dur": 2.900,
    "args": {
      "External id": 536124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622381944.088, "dur": 1.565,
    "args": {
      "External id": 536125,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622381952.705, "dur": 2662.728,
    "args": {
      "External id": 536126,"Record function id": 0, "Ev Idx": 1597
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.19)", "pid": 4183438, "tid": 31367,
    "ts": 679622381983.132, "dur": 992.802,
    "args": {
      "External id": 536127,"Record function id": 0, "Ev Idx": 1598
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.18", "pid": 4183438, "tid": 31367,
    "ts": 679622382006.033, "dur": 960.732,
    "args": {
      "External id": 536128,"Record function id": 0, "Ev Idx": 1599
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.18)", "pid": 4183438, "tid": 31367,
    "ts": 679622382019.155, "dur": 932.411,
    "args": {
      "External id": 536129,"Record function id": 0, "Ev Idx": 1600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622382092.943, "dur": 24.669,
    "args": {
      "External id": 536130,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622382135.080, "dur": 34.983,
    "args": {
      "External id": 536131,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 1602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382139.428, "dur": 1.801,
    "args": {
      "External id": 536132,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382143.107, "dur": 2.625,
    "args": {
      "External id": 536133,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382149.391, "dur": 0.495,
    "args": {
      "External id": 536134,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382151.662, "dur": 0.344,
    "args": {
      "External id": 536135,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382153.539, "dur": 0.478,
    "args": {
      "External id": 536136,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382157.324, "dur": 0.612,
    "args": {
      "External id": 536137,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382159.422, "dur": 0.520,
    "args": {
      "External id": 536138,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382161.542, "dur": 1.751,
    "args": {
      "External id": 536139,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382165.081, "dur": 0.706,
    "args": {
      "External id": 536140,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622382186.112, "dur": 33.935,
    "args": {
      "External id": 536141,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 1612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622382258.158, "dur": 109.438,
    "args": {
      "External id": 536142,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 1613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622382268.708, "dur": 6.915,
    "args": {
      "External id": 536143,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622382280.863, "dur": 11.779,
    "args": {
      "External id": 536144,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 1615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622382286.876, "dur": 5.395,
    "args": {
      "External id": 536145,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 1616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382290.525, "dur": 0.736,
    "args": {
      "External id": 536146,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 1617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622382299.337, "dur": 28.032,
    "args": {
      "External id": 536147,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 1618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382301.637, "dur": 0.566,
    "args": {
      "External id": 536148,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382303.998, "dur": 1.804,
    "args": {
      "External id": 536149,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382307.601, "dur": 0.536,
    "args": {
      "External id": 536150,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382309.311, "dur": 0.374,
    "args": {
      "External id": 536151,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382312.245, "dur": 0.360,
    "args": {
      "External id": 536152,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382314.032, "dur": 0.481,
    "args": {
      "External id": 536153,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382315.889, "dur": 2.279,
    "args": {
      "External id": 536154,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382321.169, "dur": 0.611,
    "args": {
      "External id": 536155,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622382323.369, "dur": 0.416,
    "args": {
      "External id": 536156,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622382338.365, "dur": 20.767,
    "args": {
      "External id": 536157,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 1628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622382422.938, "dur": 421.788,
    "args": {
      "External id": 536158,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 1629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622382454.217, "dur": 384.351,
    "args": {
      "External id": 536159,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1630, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622382464.454, "dur": 367.472,
    "args": {
      "External id": 536160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 1631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622382868.998, "dur": 2.305,
    "args": {
      "External id": 536161,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1632, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622382984.168, "dur": 1611.037,
    "args": {
      "External id": 536162,"Sequence number": 6319456, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1633
    }
  },
  {
    "ph": "f", "id": 172, "pid": 4183438, "tid": 31367, "ts": 679622382984.168,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622383127.762, "dur": 108.249,
    "args": {
      "External id": 536163,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 1634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622383281.134, "dur": 38.697,
    "args": {
      "External id": 536164,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 1635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622383340.790, "dur": 40.013,
    "args": {
      "External id": 536165,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 1636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622383389.678, "dur": 26.377,
    "args": {
      "External id": 536166,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 1637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622383421.893, "dur": 32.144,
    "args": {
      "External id": 536167,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 1638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622383462.840, "dur": 21.034,
    "args": {
      "External id": 536168,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 1639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622383490.953, "dur": 29.413,
    "args": {
      "External id": 536169,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 1640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622383546.863, "dur": 22.633,
    "args": {
      "External id": 536170,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 1641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622383591.285, "dur": 30.694,
    "args": {
      "External id": 536171,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622383643.290, "dur": 58.508,
    "args": {
      "External id": 536172,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 1643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622383721.994, "dur": 16.153,
    "args": {
      "External id": 536173,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 1644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622383751.193, "dur": 42.206,
    "args": {
      "External id": 536174,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622383797.586, "dur": 30.430,
    "args": {
      "External id": 536175,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622383855.838, "dur": 170.293,
    "args": {
      "External id": 536176,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622383933.296, "dur": 5.748,
    "args": {
      "External id": 536177,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622383941.075, "dur": 4.876,
    "args": {
      "External id": 536178,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622384059.907, "dur": 25.501,
    "args": {
      "External id": 536179,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622384113.847, "dur": 18.430,
    "args": {
      "External id": 536180,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622384143.918, "dur": 50.022,
    "args": {
      "External id": 536181,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622384200.081, "dur": 33.431,
    "args": {
      "External id": 536182,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622384240.935, "dur": 28.156,
    "args": {
      "External id": 536183,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622384279.161, "dur": 29.004,
    "args": {
      "External id": 536184,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622384313.935, "dur": 26.395,
    "args": {
      "External id": 536185,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622384347.320, "dur": 27.588,
    "args": {
      "External id": 536186,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622384396.001, "dur": 24.609,
    "args": {
      "External id": 536187,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 1658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622384440.719, "dur": 27.256,
    "args": {
      "External id": 536188,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622384485.213, "dur": 17.166,
    "args": {
      "External id": 536189,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 1660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622384517.515, "dur": 15.766,
    "args": {
      "External id": 536190,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 1661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622384547.782, "dur": 18.837,
    "args": {
      "External id": 536191,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 1662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384640.898, "dur": 55.583,
    "args": {
      "External id": 536192,"Record function id": 0, "Ev Idx": 1663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384644.403, "dur": 50.367,
    "args": {
      "External id": 536193,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384648.606, "dur": 44.374,
    "args": {
      "External id": 536194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384650.416, "dur": 41.886,
    "args": {
      "External id": 536195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384703.510, "dur": 6.717,
    "args": {
      "External id": 536196,"Record function id": 0, "Ev Idx": 1667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384705.517, "dur": 4.283,
    "args": {
      "External id": 536197,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384706.917, "dur": 2.319,
    "args": {
      "External id": 536198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384707.799, "dur": 1.351,
    "args": {
      "External id": 536199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384713.635, "dur": 3.781,
    "args": {
      "External id": 536200,"Record function id": 0, "Ev Idx": 1671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384714.699, "dur": 2.304,
    "args": {
      "External id": 536201,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384715.244, "dur": 1.288,
    "args": {
      "External id": 536202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384715.707, "dur": 0.732,
    "args": {
      "External id": 536203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384720.610, "dur": 5.784,
    "args": {
      "External id": 536204,"Record function id": 0, "Ev Idx": 1675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384722.047, "dur": 3.941,
    "args": {
      "External id": 536205,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384722.879, "dur": 2.651,
    "args": {
      "External id": 536206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384723.199, "dur": 2.262,
    "args": {
      "External id": 536207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384729.526, "dur": 5.647,
    "args": {
      "External id": 536208,"Record function id": 0, "Ev Idx": 1679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384730.618, "dur": 4.169,
    "args": {
      "External id": 536209,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384731.147, "dur": 3.165,
    "args": {
      "External id": 536210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384733.480, "dur": 0.723,
    "args": {
      "External id": 536211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384738.226, "dur": 3.384,
    "args": {
      "External id": 536212,"Record function id": 0, "Ev Idx": 1683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384739.326, "dur": 1.897,
    "args": {
      "External id": 536213,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384739.797, "dur": 0.950,
    "args": {
      "External id": 536214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384740.129, "dur": 0.534,
    "args": {
      "External id": 536215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384744.809, "dur": 3.980,
    "args": {
      "External id": 536216,"Record function id": 0, "Ev Idx": 1687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384746.017, "dur": 2.343,
    "args": {
      "External id": 536217,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384746.520, "dur": 1.265,
    "args": {
      "External id": 536218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384747.130, "dur": 0.591,
    "args": {
      "External id": 536219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384751.893, "dur": 4.777,
    "args": {
      "External id": 536220,"Record function id": 0, "Ev Idx": 1691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384753.181, "dur": 3.088,
    "args": {
      "External id": 536221,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384754.049, "dur": 1.813,
    "args": {
      "External id": 536222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384754.992, "dur": 0.801,
    "args": {
      "External id": 536223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384760.712, "dur": 6.546,
    "args": {
      "External id": 536224,"Record function id": 0, "Ev Idx": 1695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622384762.017, "dur": 4.824,
    "args": {
      "External id": 536225,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384762.867, "dur": 3.495,
    "args": {
      "External id": 536226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622384765.475, "dur": 0.791,
    "args": {
      "External id": 536227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622384771.411, "dur": 57874.891,
    "args": {
      "External id": 536228,"Record function id": 0, "Sequence number": 6319455, "Fwd thread id": 1, "Ev Idx": 1699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622384772.776, "dur": 57865.296,
    "args": {
      "External id": 536229,"Sequence number": 6319455, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1700
    }
  },
  {
    "ph": "f", "id": 173, "pid": 4183438, "tid": 31367, "ts": 679622384772.776,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.19)", "pid": 4183438, "tid": 31367,
    "ts": 679622384804.413, "dur": 40.877,
    "args": {
      "External id": 536230,"Record function id": 0, "Ev Idx": 1701
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.19)", "pid": 4183438, "tid": 31367,
    "ts": 679622384853.550, "dur": 71.133,
    "args": {
      "External id": 536231,"Record function id": 0, "Ev Idx": 1702
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.19)", "pid": 4183438, "tid": 31367,
    "ts": 679622384930.597, "dur": 57699.802,
    "args": {
      "External id": 536232,"Record function id": 0, "Ev Idx": 1703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622384985.157, "dur": 7.613,
    "args": {
      "External id": 536233,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622385002.581, "dur": 4.829,
    "args": {
      "External id": 536234,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 1705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622385025.035, "dur": 56798.138,
    "args": {
      "External id": 536235,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 1706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622385040.138, "dur": 56770.144,
    "args": {
      "External id": 536236,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 1707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622385080.242, "dur": 6.590,
    "args": {
      "External id": 536237,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622385091.366, "dur": 56669.297,
    "args": {
      "External id": 536238,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 1709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622385093.625, "dur": 56665.976,
    "args": {
      "External id": 536239,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 1710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622385114.612, "dur": 6.701,
    "args": {
      "External id": 536240,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622385123.436, "dur": 56630.150,
    "args": {
      "External id": 536241,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 1712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622441933.339, "dur": 13.387,
    "args": {
      "External id": 536242,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 1713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622441936.999, "dur": 9.225,
    "args": {
      "External id": 536243,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622441987.924, "dur": 326.902,
    "args": {
      "External id": 536244,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 1715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622442020.954, "dur": 288.509,
    "args": {
      "External id": 536245,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1716, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622442034.890, "dur": 268.381,
    "args": {
      "External id": 536246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 1717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622442337.574, "dur": 2.258,
    "args": {
      "External id": 536247,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1718, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442401.646, "dur": 6.957,
    "args": {
      "External id": 536248,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442461.153, "dur": 1.349,
    "args": {
      "External id": 536249,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442481.541, "dur": 1.264,
    "args": {
      "External id": 536250,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442497.009, "dur": 0.752,
    "args": {
      "External id": 536251,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442511.197, "dur": 1.083,
    "args": {
      "External id": 536252,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442523.365, "dur": 0.793,
    "args": {
      "External id": 536253,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442536.424, "dur": 1.137,
    "args": {
      "External id": 536254,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442552.435, "dur": 1.873,
    "args": {
      "External id": 536255,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442565.682, "dur": 1.071,
    "args": {
      "External id": 536256,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622442698.960, "dur": 3379.982,
    "args": {
      "External id": 536257,"Record function id": 0, "Ev Idx": 1728
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.18)", "pid": 4183438, "tid": 31367,
    "ts": 679622442724.663, "dur": 1084.819,
    "args": {
      "External id": 536258,"Record function id": 0, "Ev Idx": 1729
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.18)", "pid": 4183438, "tid": 31367,
    "ts": 679622442739.152, "dur": 319.569,
    "args": {
      "External id": 536259,"Record function id": 0, "Ev Idx": 1730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442821.715, "dur": 4.732,
    "args": {
      "External id": 536260,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 1731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442829.797, "dur": 0.965,
    "args": {
      "External id": 536261,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 1732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442832.845, "dur": 0.983,
    "args": {
      "External id": 536262,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442835.505, "dur": 2.728,
    "args": {
      "External id": 536263,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442841.832, "dur": 1.035,
    "args": {
      "External id": 536264,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442844.400, "dur": 0.903,
    "args": {
      "External id": 536265,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442847.070, "dur": 1.994,
    "args": {
      "External id": 536266,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 1737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442850.693, "dur": 1.022,
    "args": {
      "External id": 536267,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 1738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442854.970, "dur": 1.004,
    "args": {
      "External id": 536268,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 1739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622442857.282, "dur": 1.207,
    "args": {
      "External id": 536269,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 1740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622442876.912, "dur": 151.548,
    "args": {
      "External id": 536270,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 1741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622442894.044, "dur": 129.909,
    "args": {
      "External id": 536271,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 1742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622442912.400, "dur": 13.195,
    "args": {
      "External id": 536272,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622442928.346, "dur": 66.692,
    "args": {
      "External id": 536273,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 1744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622442931.111, "dur": 63.589,
    "args": {
      "External id": 536274,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 1745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622442935.404, "dur": 7.487,
    "args": {
      "External id": 536275,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622442946.278, "dur": 47.766,
    "args": {
      "External id": 536276,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 1747
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.17", "pid": 4183438, "tid": 31367,
    "ts": 679622443159.664, "dur": 642.438,
    "args": {
      "External id": 536277,"Record function id": 0, "Ev Idx": 1748
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.17)", "pid": 4183438, "tid": 31367,
    "ts": 679622443177.255, "dur": 612.067,
    "args": {
      "External id": 536278,"Record function id": 0, "Ev Idx": 1749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622443232.917, "dur": 5.553,
    "args": {
      "External id": 536279,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622443253.800, "dur": 33.762,
    "args": {
      "External id": 536280,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 1751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443258.819, "dur": 1.782,
    "args": {
      "External id": 536281,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443263.031, "dur": 0.537,
    "args": {
      "External id": 536282,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443265.428, "dur": 0.628,
    "args": {
      "External id": 536283,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443268.413, "dur": 0.513,
    "args": {
      "External id": 536284,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443270.464, "dur": 0.503,
    "args": {
      "External id": 536285,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443272.433, "dur": 1.743,
    "args": {
      "External id": 536286,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443275.453, "dur": 2.044,
    "args": {
      "External id": 536287,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443278.897, "dur": 0.291,
    "args": {
      "External id": 536288,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443282.075, "dur": 0.371,
    "args": {
      "External id": 536289,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622443300.881, "dur": 33.377,
    "args": {
      "External id": 536290,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 1761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622443366.912, "dur": 105.675,
    "args": {
      "External id": 536291,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 1762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622443376.755, "dur": 3.907,
    "args": {
      "External id": 536292,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622443385.856, "dur": 10.240,
    "args": {
      "External id": 536293,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 1764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622443390.443, "dur": 5.266,
    "args": {
      "External id": 536294,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 1765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443394.060, "dur": 0.526,
    "args": {
      "External id": 536295,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 1766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622443402.831, "dur": 29.539,
    "args": {
      "External id": 536296,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 1767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443405.288, "dur": 0.542,
    "args": {
      "External id": 536297,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443407.455, "dur": 1.874,
    "args": {
      "External id": 536298,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443410.945, "dur": 0.584,
    "args": {
      "External id": 536299,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443412.796, "dur": 2.106,
    "args": {
      "External id": 536300,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443418.506, "dur": 0.423,
    "args": {
      "External id": 536301,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443420.268, "dur": 0.529,
    "args": {
      "External id": 536302,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443422.239, "dur": 0.301,
    "args": {
      "External id": 536303,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443425.391, "dur": 0.512,
    "args": {
      "External id": 536304,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622443427.398, "dur": 0.386,
    "args": {
      "External id": 536305,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622443445.188, "dur": 19.266,
    "args": {
      "External id": 536306,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 1777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622443544.526, "dur": 167.302,
    "args": {
      "External id": 536307,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 1778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622443574.321, "dur": 133.388,
    "args": {
      "External id": 536308,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1779, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622443587.441, "dur": 115.052,
    "args": {
      "External id": 536309,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 1780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622443730.251, "dur": 2.054,
    "args": {
      "External id": 536310,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1781, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622443816.750, "dur": 2239.960,
    "args": {
      "External id": 536311,"Sequence number": 6319454, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1782
    }
  },
  {
    "ph": "f", "id": 174, "pid": 4183438, "tid": 31367, "ts": 679622443816.750,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622443931.590, "dur": 104.407,
    "args": {
      "External id": 536312,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 1783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622444077.969, "dur": 55.755,
    "args": {
      "External id": 536313,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 1784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622444153.805, "dur": 45.019,
    "args": {
      "External id": 536314,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 1785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622444208.648, "dur": 25.297,
    "args": {
      "External id": 536315,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 1786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622444243.189, "dur": 32.651,
    "args": {
      "External id": 536316,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 1787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622444282.289, "dur": 20.671,
    "args": {
      "External id": 536317,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 1788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622444309.894, "dur": 28.891,
    "args": {
      "External id": 536318,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 1789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622444366.766, "dur": 22.513,
    "args": {
      "External id": 536319,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 1790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622444409.300, "dur": 28.173,
    "args": {
      "External id": 536320,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622444459.403, "dur": 17.853,
    "args": {
      "External id": 536321,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 1792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622444492.496, "dur": 15.825,
    "args": {
      "External id": 536322,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 1793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622444515.231, "dur": 36.237,
    "args": {
      "External id": 536323,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622444554.675, "dur": 29.686,
    "args": {
      "External id": 536324,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622444611.863, "dur": 223.362,
    "args": {
      "External id": 536325,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622444736.317, "dur": 8.049,
    "args": {
      "External id": 536326,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622444746.510, "dur": 2.469,
    "args": {
      "External id": 536327,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622444873.152, "dur": 29.970,
    "args": {
      "External id": 536328,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622444917.991, "dur": 14.305,
    "args": {
      "External id": 536329,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622444940.224, "dur": 54.168,
    "args": {
      "External id": 536330,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622445002.504, "dur": 43.530,
    "args": {
      "External id": 536331,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622445061.321, "dur": 58.293,
    "args": {
      "External id": 536332,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622445129.513, "dur": 43.202,
    "args": {
      "External id": 536333,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622445185.576, "dur": 64.500,
    "args": {
      "External id": 536334,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622445267.433, "dur": 50.197,
    "args": {
      "External id": 536335,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622445356.371, "dur": 258.678,
    "args": {
      "External id": 536336,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 1807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622445849.642, "dur": 43.152,
    "args": {
      "External id": 536337,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622445918.493, "dur": 27.313,
    "args": {
      "External id": 536338,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 1809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622445970.113, "dur": 15.160,
    "args": {
      "External id": 536339,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 1810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622446005.419, "dur": 16.760,
    "args": {
      "External id": 536340,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 1811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446116.213, "dur": 18.814,
    "args": {
      "External id": 536341,"Record function id": 0, "Ev Idx": 1812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446121.173, "dur": 12.664,
    "args": {
      "External id": 536342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446125.808, "dur": 6.778,
    "args": {
      "External id": 536343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446127.667, "dur": 4.672,
    "args": {
      "External id": 536344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446140.234, "dur": 5.446,
    "args": {
      "External id": 536345,"Record function id": 0, "Ev Idx": 1816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446141.956, "dur": 3.266,
    "args": {
      "External id": 536346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446142.783, "dur": 1.995,
    "args": {
      "External id": 536347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446143.636, "dur": 1.019,
    "args": {
      "External id": 536348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446148.898, "dur": 7.005,
    "args": {
      "External id": 536349,"Record function id": 0, "Ev Idx": 1820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446150.422, "dur": 5.068,
    "args": {
      "External id": 536350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446151.318, "dur": 3.648,
    "args": {
      "External id": 536351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446151.909, "dur": 2.952,
    "args": {
      "External id": 536352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446159.458, "dur": 4.764,
    "args": {
      "External id": 536353,"Record function id": 0, "Ev Idx": 1824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446160.801, "dur": 3.023,
    "args": {
      "External id": 536354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446161.497, "dur": 1.913,
    "args": {
      "External id": 536355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446162.211, "dur": 1.135,
    "args": {
      "External id": 536356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446167.273, "dur": 3.995,
    "args": {
      "External id": 536357,"Record function id": 0, "Ev Idx": 1828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446168.529, "dur": 2.353,
    "args": {
      "External id": 536358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446169.233, "dur": 1.214,
    "args": {
      "External id": 536359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446169.544, "dur": 0.838,
    "args": {
      "External id": 536360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446174.288, "dur": 4.678,
    "args": {
      "External id": 536361,"Record function id": 0, "Ev Idx": 1832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446175.908, "dur": 2.675,
    "args": {
      "External id": 536362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446176.573, "dur": 1.562,
    "args": {
      "External id": 536363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446177.415, "dur": 0.656,
    "args": {
      "External id": 536364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446182.237, "dur": 4.111,
    "args": {
      "External id": 536365,"Record function id": 0, "Ev Idx": 1836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446183.501, "dur": 2.433,
    "args": {
      "External id": 536366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446184.184, "dur": 1.302,
    "args": {
      "External id": 536367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446184.525, "dur": 0.865,
    "args": {
      "External id": 536368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446189.328, "dur": 5.935,
    "args": {
      "External id": 536369,"Record function id": 0, "Ev Idx": 1840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446190.521, "dur": 4.344,
    "args": {
      "External id": 536370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446191.069, "dur": 3.372,
    "args": {
      "External id": 536371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446193.571, "dur": 0.808,
    "args": {
      "External id": 536372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446198.231, "dur": 4.526,
    "args": {
      "External id": 536373,"Record function id": 0, "Ev Idx": 1844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622446199.886, "dur": 2.477,
    "args": {
      "External id": 536374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446200.477, "dur": 1.462,
    "args": {
      "External id": 536375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622446200.914, "dur": 0.961,
    "args": {
      "External id": 536376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622446206.591, "dur": 15270.242,
    "args": {
      "External id": 536377,"Record function id": 0, "Sequence number": 6319453, "Fwd thread id": 1, "Ev Idx": 1848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622446207.948, "dur": 15260.722,
    "args": {
      "External id": 536378,"Sequence number": 6319453, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1849
    }
  },
  {
    "ph": "f", "id": 175, "pid": 4183438, "tid": 31367, "ts": 679622446207.948,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.18)", "pid": 4183438, "tid": 31367,
    "ts": 679622446241.013, "dur": 40.951,
    "args": {
      "External id": 536379,"Record function id": 0, "Ev Idx": 1850
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.18)", "pid": 4183438, "tid": 31367,
    "ts": 679622446290.089, "dur": 65.957,
    "args": {
      "External id": 536380,"Record function id": 0, "Ev Idx": 1851
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.18)", "pid": 4183438, "tid": 31367,
    "ts": 679622446364.876, "dur": 15095.710,
    "args": {
      "External id": 536381,"Record function id": 0, "Ev Idx": 1852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622446463.477, "dur": 7.486,
    "args": {
      "External id": 536382,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622446480.273, "dur": 6.508,
    "args": {
      "External id": 536383,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 1854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622446501.771, "dur": 14214.582,
    "args": {
      "External id": 536384,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 1855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622446519.089, "dur": 14185.999,
    "args": {
      "External id": 536385,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 1856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622446548.419, "dur": 14.206,
    "args": {
      "External id": 536386,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622446570.266, "dur": 14068.805,
    "args": {
      "External id": 536387,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 1858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622446572.983, "dur": 14065.092,
    "args": {
      "External id": 536388,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 1859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622446578.438, "dur": 4.777,
    "args": {
      "External id": 536389,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622446586.659, "dur": 14046.336,
    "args": {
      "External id": 536390,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 1861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622460824.172, "dur": 13.025,
    "args": {
      "External id": 536391,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 1862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622460828.027, "dur": 8.821,
    "args": {
      "External id": 536392,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622460868.348, "dur": 279.555,
    "args": {
      "External id": 536393,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 1864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622460900.730, "dur": 242.218,
    "args": {
      "External id": 536394,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1865, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622460914.067, "dur": 223.657,
    "args": {
      "External id": 536395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 1866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622461169.767, "dur": 2.588,
    "args": {
      "External id": 536396,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1867, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461231.608, "dur": 7.021,
    "args": {
      "External id": 536397,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461286.947, "dur": 1.610,
    "args": {
      "External id": 536398,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461306.297, "dur": 3.596,
    "args": {
      "External id": 536399,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461321.881, "dur": 1.078,
    "args": {
      "External id": 536400,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461335.581, "dur": 0.977,
    "args": {
      "External id": 536401,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461348.989, "dur": 0.800,
    "args": {
      "External id": 536402,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461362.266, "dur": 3.147,
    "args": {
      "External id": 536403,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461378.022, "dur": 2.057,
    "args": {
      "External id": 536404,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461393.999, "dur": 0.894,
    "args": {
      "External id": 536405,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622461491.030, "dur": 3269.656,
    "args": {
      "External id": 536406,"Record function id": 0, "Ev Idx": 1877
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.17)", "pid": 4183438, "tid": 31367,
    "ts": 679622461510.217, "dur": 1030.628,
    "args": {
      "External id": 536407,"Record function id": 0, "Ev Idx": 1878
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.17)", "pid": 4183438, "tid": 31367,
    "ts": 679622461525.359, "dur": 350.679,
    "args": {
      "External id": 536408,"Record function id": 0, "Ev Idx": 1879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461602.154, "dur": 4.210,
    "args": {
      "External id": 536409,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 1880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461609.278, "dur": 0.943,
    "args": {
      "External id": 536410,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 1881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461612.057, "dur": 2.841,
    "args": {
      "External id": 536411,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461618.305, "dur": 0.889,
    "args": {
      "External id": 536412,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461620.595, "dur": 1.098,
    "args": {
      "External id": 536413,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461623.346, "dur": 0.957,
    "args": {
      "External id": 536414,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 1885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461625.812, "dur": 1.934,
    "args": {
      "External id": 536415,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 1886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461631.078, "dur": 0.658,
    "args": {
      "External id": 536416,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 1887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461633.213, "dur": 0.960,
    "args": {
      "External id": 536417,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 1888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622461635.778, "dur": 0.750,
    "args": {
      "External id": 536418,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 1889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622461690.942, "dur": 153.201,
    "args": {
      "External id": 536419,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 1890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622461708.848, "dur": 130.215,
    "args": {
      "External id": 536420,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 1891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622461728.162, "dur": 15.937,
    "args": {
      "External id": 536421,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622461747.280, "dur": 65.468,
    "args": {
      "External id": 536422,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 1893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622461750.049, "dur": 62.371,
    "args": {
      "External id": 536423,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 1894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622461755.269, "dur": 6.616,
    "args": {
      "External id": 536424,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622461763.501, "dur": 48.205,
    "args": {
      "External id": 536425,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 1896
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.16", "pid": 4183438, "tid": 31367,
    "ts": 679622461959.117, "dur": 574.757,
    "args": {
      "External id": 536426,"Record function id": 0, "Ev Idx": 1897
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.16)", "pid": 4183438, "tid": 31367,
    "ts": 679622461975.757, "dur": 545.172,
    "args": {
      "External id": 536427,"Record function id": 0, "Ev Idx": 1898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622462030.646, "dur": 4.973,
    "args": {
      "External id": 536428,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622462050.835, "dur": 32.893,
    "args": {
      "External id": 536429,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 1900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462055.266, "dur": 1.677,
    "args": {
      "External id": 536430,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462058.980, "dur": 1.728,
    "args": {
      "External id": 536431,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462062.104, "dur": 0.540,
    "args": {
      "External id": 536432,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462064.175, "dur": 0.581,
    "args": {
      "External id": 536433,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462067.822, "dur": 0.278,
    "args": {
      "External id": 536434,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462069.340, "dur": 2.153,
    "args": {
      "External id": 536435,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462072.815, "dur": 0.568,
    "args": {
      "External id": 536436,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462076.038, "dur": 0.548,
    "args": {
      "External id": 536437,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462078.108, "dur": 0.659,
    "args": {
      "External id": 536438,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622462093.968, "dur": 49.354,
    "args": {
      "External id": 536439,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 1910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622462179.471, "dur": 110.744,
    "args": {
      "External id": 536440,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 1911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622462189.486, "dur": 4.234,
    "args": {
      "External id": 536441,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622462199.186, "dur": 10.730,
    "args": {
      "External id": 536442,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 1913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622462203.244, "dur": 6.242,
    "args": {
      "External id": 536443,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 1914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462207.223, "dur": 0.773,
    "args": {
      "External id": 536444,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 1915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622462217.328, "dur": 32.799,
    "args": {
      "External id": 536445,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 1916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462219.443, "dur": 0.587,
    "args": {
      "External id": 536446,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462223.036, "dur": 0.292,
    "args": {
      "External id": 536447,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462229.484, "dur": 1.859,
    "args": {
      "External id": 536448,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462232.886, "dur": 1.236,
    "args": {
      "External id": 536449,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462235.809, "dur": 0.600,
    "args": {
      "External id": 536450,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462237.692, "dur": 0.520,
    "args": {
      "External id": 536451,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462241.414, "dur": 0.325,
    "args": {
      "External id": 536452,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462243.474, "dur": 0.344,
    "args": {
      "External id": 536453,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622462245.083, "dur": 0.521,
    "args": {
      "External id": 536454,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 1925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622462261.117, "dur": 20.986,
    "args": {
      "External id": 536455,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 1926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622462334.979, "dur": 117.739,
    "args": {
      "External id": 536456,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 1927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622462364.852, "dur": 84.396,
    "args": {
      "External id": 536457,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1928, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622462374.641, "dur": 70.601,
    "args": {
      "External id": 536458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 1929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622462469.608, "dur": 1.991,
    "args": {
      "External id": 536459,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1930, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622462547.747, "dur": 2189.608,
    "args": {
      "External id": 536460,"Sequence number": 6319452, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1931
    }
  },
  {
    "ph": "f", "id": 176, "pid": 4183438, "tid": 31367, "ts": 679622462547.747,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622462698.518, "dur": 102.963,
    "args": {
      "External id": 536461,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 1932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622462845.379, "dur": 35.873,
    "args": {
      "External id": 536462,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 1933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622462899.120, "dur": 39.402,
    "args": {
      "External id": 536463,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 1934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622462950.176, "dur": 24.911,
    "args": {
      "External id": 536464,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 1935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622462982.085, "dur": 34.027,
    "args": {
      "External id": 536465,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 1936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463022.840, "dur": 20.746,
    "args": {
      "External id": 536466,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 1937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463056.059, "dur": 31.978,
    "args": {
      "External id": 536467,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 1938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622463140.469, "dur": 26.547,
    "args": {
      "External id": 536468,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 1939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622463188.582, "dur": 27.213,
    "args": {
      "External id": 536469,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622463242.853, "dur": 18.538,
    "args": {
      "External id": 536470,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 1941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622463275.329, "dur": 13.464,
    "args": {
      "External id": 536471,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 1942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463301.235, "dur": 41.015,
    "args": {
      "External id": 536472,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463345.707, "dur": 30.661,
    "args": {
      "External id": 536473,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622463404.884, "dur": 176.099,
    "args": {
      "External id": 536474,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622463483.809, "dur": 7.962,
    "args": {
      "External id": 536475,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622463493.949, "dur": 5.606,
    "args": {
      "External id": 536476,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622463610.762, "dur": 23.113,
    "args": {
      "External id": 536477,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622463647.153, "dur": 54.801,
    "args": {
      "External id": 536478,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463713.439, "dur": 57.103,
    "args": {
      "External id": 536479,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463776.775, "dur": 35.233,
    "args": {
      "External id": 536480,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463823.409, "dur": 39.791,
    "args": {
      "External id": 536481,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463875.691, "dur": 41.515,
    "args": {
      "External id": 536482,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463930.275, "dur": 39.099,
    "args": {
      "External id": 536483,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 1954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622463979.077, "dur": 60.308,
    "args": {
      "External id": 536484,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 1955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622464085.826, "dur": 56.674,
    "args": {
      "External id": 536485,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 1956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622464173.954, "dur": 199.351,
    "args": {
      "External id": 536486,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622464561.740, "dur": 25.845,
    "args": {
      "External id": 536487,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 1958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622464610.940, "dur": 20.073,
    "args": {
      "External id": 536488,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 1959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622464649.827, "dur": 51.492,
    "args": {
      "External id": 536489,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 1960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464784.667, "dur": 20.039,
    "args": {
      "External id": 536490,"Record function id": 0, "Ev Idx": 1961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464788.485, "dur": 11.955,
    "args": {
      "External id": 536491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464793.338, "dur": 6.157,
    "args": {
      "External id": 536492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464795.135, "dur": 4.280,
    "args": {
      "External id": 536493,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464808.805, "dur": 6.810,
    "args": {
      "External id": 536494,"Record function id": 0, "Ev Idx": 1965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464810.188, "dur": 4.970,
    "args": {
      "External id": 536495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464811.043, "dur": 3.597,
    "args": {
      "External id": 536496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464811.604, "dur": 2.919,
    "args": {
      "External id": 536497,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464819.173, "dur": 5.665,
    "args": {
      "External id": 536498,"Record function id": 0, "Ev Idx": 1969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464820.531, "dur": 3.876,
    "args": {
      "External id": 536499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464821.569, "dur": 2.399,
    "args": {
      "External id": 536500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464822.726, "dur": 1.102,
    "args": {
      "External id": 536501,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 1972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464827.901, "dur": 4.069,
    "args": {
      "External id": 536502,"Record function id": 0, "Ev Idx": 1973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464829.224, "dur": 2.350,
    "args": {
      "External id": 536503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464829.733, "dur": 1.360,
    "args": {
      "External id": 536504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464830.466, "dur": 0.561,
    "args": {
      "External id": 536505,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 1976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464835.209, "dur": 3.819,
    "args": {
      "External id": 536506,"Record function id": 0, "Ev Idx": 1977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464836.334, "dur": 2.263,
    "args": {
      "External id": 536507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464836.773, "dur": 1.411,
    "args": {
      "External id": 536508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464837.255, "dur": 0.833,
    "args": {
      "External id": 536509,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464842.102, "dur": 4.131,
    "args": {
      "External id": 536510,"Record function id": 0, "Ev Idx": 1981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464843.373, "dur": 2.475,
    "args": {
      "External id": 536511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464844.136, "dur": 1.308,
    "args": {
      "External id": 536512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464844.563, "dur": 0.786,
    "args": {
      "External id": 536513,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464849.582, "dur": 4.725,
    "args": {
      "External id": 536514,"Record function id": 0, "Ev Idx": 1985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464851.475, "dur": 2.445,
    "args": {
      "External id": 536515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464852.022, "dur": 1.494,
    "args": {
      "External id": 536516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464852.475, "dur": 0.947,
    "args": {
      "External id": 536517,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464857.345, "dur": 9.774,
    "args": {
      "External id": 536518,"Record function id": 0, "Ev Idx": 1989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464858.943, "dur": 7.761,
    "args": {
      "External id": 536519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464862.644, "dur": 3.661,
    "args": {
      "External id": 536520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464865.324, "dur": 0.888,
    "args": {
      "External id": 536521,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 1992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464870.111, "dur": 6.304,
    "args": {
      "External id": 536522,"Record function id": 0, "Ev Idx": 1993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622464871.378, "dur": 4.646,
    "args": {
      "External id": 536523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464872.272, "dur": 3.346,
    "args": {
      "External id": 536524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622464872.809, "dur": 2.747,
    "args": {
      "External id": 536525,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 1996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622464880.068, "dur": 15032.467,
    "args": {
      "External id": 536526,"Record function id": 0, "Sequence number": 6319451, "Fwd thread id": 1, "Ev Idx": 1997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622464881.632, "dur": 15021.573,
    "args": {
      "External id": 536527,"Sequence number": 6319451, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 1998
    }
  },
  {
    "ph": "f", "id": 177, "pid": 4183438, "tid": 31367, "ts": 679622464881.632,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.17)", "pid": 4183438, "tid": 31367,
    "ts": 679622464915.259, "dur": 45.513,
    "args": {
      "External id": 536528,"Record function id": 0, "Ev Idx": 1999
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.17)", "pid": 4183438, "tid": 31367,
    "ts": 679622464968.856, "dur": 68.020,
    "args": {
      "External id": 536529,"Record function id": 0, "Ev Idx": 2000
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.17)", "pid": 4183438, "tid": 31367,
    "ts": 679622465045.131, "dur": 14849.987,
    "args": {
      "External id": 536530,"Record function id": 0, "Ev Idx": 2001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622465166.874, "dur": 14.246,
    "args": {
      "External id": 536531,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622465200.496, "dur": 8.021,
    "args": {
      "External id": 536532,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622465227.404, "dur": 13910.319,
    "args": {
      "External id": 536533,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622465244.049, "dur": 13882.540,
    "args": {
      "External id": 536534,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622465273.914, "dur": 15.702,
    "args": {
      "External id": 536535,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622465293.620, "dur": 13781.939,
    "args": {
      "External id": 536536,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 2007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622465295.832, "dur": 13778.730,
    "args": {
      "External id": 536537,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 2008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622465300.321, "dur": 6.890,
    "args": {
      "External id": 536538,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622465310.835, "dur": 13759.244,
    "args": {
      "External id": 536539,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 2010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622479242.556, "dur": 12.091,
    "args": {
      "External id": 536540,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 2011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622479246.267, "dur": 8.013,
    "args": {
      "External id": 536541,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622479284.158, "dur": 265.035,
    "args": {
      "External id": 536542,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 2013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622479315.789, "dur": 229.138,
    "args": {
      "External id": 536543,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2014, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622479328.117, "dur": 211.040,
    "args": {
      "External id": 536544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 2015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622479569.021, "dur": 2.659,
    "args": {
      "External id": 536545,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2016, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479626.439, "dur": 6.733,
    "args": {
      "External id": 536546,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479721.193, "dur": 3.919,
    "args": {
      "External id": 536547,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479744.863, "dur": 1.258,
    "args": {
      "External id": 536548,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479759.390, "dur": 1.157,
    "args": {
      "External id": 536549,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479776.672, "dur": 1.311,
    "args": {
      "External id": 536550,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479789.496, "dur": 2.688,
    "args": {
      "External id": 536551,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479804.907, "dur": 0.767,
    "args": {
      "External id": 536552,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479817.379, "dur": 2.437,
    "args": {
      "External id": 536553,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622479830.310, "dur": 1.080,
    "args": {
      "External id": 536554,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622479928.405, "dur": 3306.742,
    "args": {
      "External id": 536555,"Record function id": 0, "Ev Idx": 2026
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.16)", "pid": 4183438, "tid": 31367,
    "ts": 679622479948.283, "dur": 1055.203,
    "args": {
      "External id": 536556,"Record function id": 0, "Ev Idx": 2027
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.16)", "pid": 4183438, "tid": 31367,
    "ts": 679622479963.326, "dur": 335.757,
    "args": {
      "External id": 536557,"Record function id": 0, "Ev Idx": 2028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480041.212, "dur": 5.971,
    "args": {
      "External id": 536558,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480050.300, "dur": 1.301,
    "args": {
      "External id": 536559,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480053.391, "dur": 1.095,
    "args": {
      "External id": 536560,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480058.249, "dur": 0.875,
    "args": {
      "External id": 536561,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480060.676, "dur": 0.884,
    "args": {
      "External id": 536562,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480062.882, "dur": 0.766,
    "args": {
      "External id": 536563,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480065.443, "dur": 2.055,
    "args": {
      "External id": 536564,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480070.063, "dur": 0.991,
    "args": {
      "External id": 536565,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480072.620, "dur": 2.878,
    "args": {
      "External id": 536566,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622480077.005, "dur": 0.782,
    "args": {
      "External id": 536567,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622480094.508, "dur": 173.880,
    "args": {
      "External id": 536568,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622480128.979, "dur": 135.010,
    "args": {
      "External id": 536569,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622480148.765, "dur": 14.414,
    "args": {
      "External id": 536570,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622480167.569, "dur": 68.011,
    "args": {
      "External id": 536571,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 2042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622480170.473, "dur": 64.749,
    "args": {
      "External id": 536572,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 2043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480176.629, "dur": 6.857,
    "args": {
      "External id": 536573,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622480185.326, "dur": 49.152,
    "args": {
      "External id": 536574,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 2045
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.15", "pid": 4183438, "tid": 31367,
    "ts": 679622480381.687, "dur": 614.384,
    "args": {
      "External id": 536575,"Record function id": 0, "Ev Idx": 2046
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.15)", "pid": 4183438, "tid": 31367,
    "ts": 679622480399.552, "dur": 584.750,
    "args": {
      "External id": 536576,"Record function id": 0, "Ev Idx": 2047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622480473.424, "dur": 4.982,
    "args": {
      "External id": 536577,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622480494.610, "dur": 31.764,
    "args": {
      "External id": 536578,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480499.600, "dur": 1.670,
    "args": {
      "External id": 536579,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480504.516, "dur": 0.654,
    "args": {
      "External id": 536580,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480506.601, "dur": 0.355,
    "args": {
      "External id": 536581,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480508.083, "dur": 3.107,
    "args": {
      "External id": 536582,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480512.601, "dur": 0.351,
    "args": {
      "External id": 536583,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480514.256, "dur": 0.300,
    "args": {
      "External id": 536584,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480517.156, "dur": 0.340,
    "args": {
      "External id": 536585,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480519.217, "dur": 0.424,
    "args": {
      "External id": 536586,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480520.923, "dur": 0.346,
    "args": {
      "External id": 536587,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622480535.386, "dur": 32.338,
    "args": {
      "External id": 536588,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622480599.256, "dur": 148.953,
    "args": {
      "External id": 536589,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 2060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622480609.225, "dur": 3.679,
    "args": {
      "External id": 536590,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622480618.038, "dur": 10.752,
    "args": {
      "External id": 536591,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622480622.189, "dur": 6.144,
    "args": {
      "External id": 536592,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 2063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480625.847, "dur": 1.180,
    "args": {
      "External id": 536593,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622480635.757, "dur": 66.196,
    "args": {
      "External id": 536594,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480637.895, "dur": 2.398,
    "args": {
      "External id": 536595,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480641.839, "dur": 0.564,
    "args": {
      "External id": 536596,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480643.913, "dur": 0.314,
    "args": {
      "External id": 536597,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480647.136, "dur": 0.307,
    "args": {
      "External id": 536598,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480648.706, "dur": 0.339,
    "args": {
      "External id": 536599,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480650.331, "dur": 38.409,
    "args": {
      "External id": 536600,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480691.798, "dur": 0.521,
    "args": {
      "External id": 536601,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480693.772, "dur": 0.261,
    "args": {
      "External id": 536602,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622480696.420, "dur": 1.863,
    "args": {
      "External id": 536603,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622480716.009, "dur": 23.690,
    "args": {
      "External id": 536604,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622480795.473, "dur": 120.949,
    "args": {
      "External id": 536605,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 2076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622480822.621, "dur": 89.936,
    "args": {
      "External id": 536606,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2077, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622480832.202, "dur": 76.159,
    "args": {
      "External id": 536607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 2078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622480932.657, "dur": 1.807,
    "args": {
      "External id": 536608,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2079, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622481010.255, "dur": 2202.989,
    "args": {
      "External id": 536609,"Sequence number": 6319450, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2080
    }
  },
  {
    "ph": "f", "id": 178, "pid": 4183438, "tid": 31367, "ts": 679622481010.255,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622481140.764, "dur": 110.141,
    "args": {
      "External id": 536610,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 2081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622481290.444, "dur": 37.843,
    "args": {
      "External id": 536611,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 2082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622481350.575, "dur": 41.571,
    "args": {
      "External id": 536612,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 2083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622481404.733, "dur": 27.397,
    "args": {
      "External id": 536613,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622481438.854, "dur": 35.776,
    "args": {
      "External id": 536614,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622481480.965, "dur": 21.993,
    "args": {
      "External id": 536615,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622481515.012, "dur": 28.870,
    "args": {
      "External id": 536616,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622481566.902, "dur": 22.649,
    "args": {
      "External id": 536617,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 2088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622481609.911, "dur": 26.914,
    "args": {
      "External id": 536618,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622481698.064, "dur": 28.370,
    "args": {
      "External id": 536619,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622481742.024, "dur": 15.770,
    "args": {
      "External id": 536620,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622481770.686, "dur": 42.891,
    "args": {
      "External id": 536621,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622481817.214, "dur": 31.047,
    "args": {
      "External id": 536622,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622481880.393, "dur": 180.008,
    "args": {
      "External id": 536623,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622481965.479, "dur": 7.450,
    "args": {
      "External id": 536624,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622481974.727, "dur": 2.633,
    "args": {
      "External id": 536625,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622482094.773, "dur": 43.737,
    "args": {
      "External id": 536626,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622482152.059, "dur": 15.246,
    "args": {
      "External id": 536627,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622482176.716, "dur": 51.079,
    "args": {
      "External id": 536628,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622482237.224, "dur": 44.427,
    "args": {
      "External id": 536629,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622482293.486, "dur": 41.400,
    "args": {
      "External id": 536630,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622482341.915, "dur": 40.468,
    "args": {
      "External id": 536631,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622482391.212, "dur": 67.423,
    "args": {
      "External id": 536632,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622482478.854, "dur": 45.966,
    "args": {
      "External id": 536633,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622482560.456, "dur": 36.216,
    "args": {
      "External id": 536634,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 2105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622482828.372, "dur": 37.511,
    "args": {
      "External id": 536635,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622483057.401, "dur": 22.169,
    "args": {
      "External id": 536636,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622483121.939, "dur": 21.651,
    "args": {
      "External id": 536637,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622483161.524, "dur": 16.687,
    "args": {
      "External id": 536638,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 2109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483258.800, "dur": 21.881,
    "args": {
      "External id": 536639,"Record function id": 0, "Ev Idx": 2110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483262.437, "dur": 17.184,
    "args": {
      "External id": 536640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483270.545, "dur": 8.150,
    "args": {
      "External id": 536641,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483272.032, "dur": 6.558,
    "args": {
      "External id": 536642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483284.549, "dur": 4.439,
    "args": {
      "External id": 536643,"Record function id": 0, "Ev Idx": 2114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483285.966, "dur": 2.571,
    "args": {
      "External id": 536644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483286.928, "dur": 1.155,
    "args": {
      "External id": 536645,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483287.281, "dur": 0.655,
    "args": {
      "External id": 536646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483292.181, "dur": 4.447,
    "args": {
      "External id": 536647,"Record function id": 0, "Ev Idx": 2118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483293.419, "dur": 2.841,
    "args": {
      "External id": 536648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483294.152, "dur": 1.704,
    "args": {
      "External id": 536649,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483294.717, "dur": 1.028,
    "args": {
      "External id": 536650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483299.828, "dur": 4.259,
    "args": {
      "External id": 536651,"Record function id": 0, "Ev Idx": 2122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483301.461, "dur": 2.207,
    "args": {
      "External id": 536652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483302.189, "dur": 1.080,
    "args": {
      "External id": 536653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483302.620, "dur": 0.543,
    "args": {
      "External id": 536654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483307.109, "dur": 4.416,
    "args": {
      "External id": 536655,"Record function id": 0, "Ev Idx": 2126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483308.121, "dur": 3.020,
    "args": {
      "External id": 536656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483308.837, "dur": 1.907,
    "args": {
      "External id": 536657,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483310.045, "dur": 0.567,
    "args": {
      "External id": 536658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483314.621, "dur": 3.284,
    "args": {
      "External id": 536659,"Record function id": 0, "Ev Idx": 2130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483315.640, "dur": 1.901,
    "args": {
      "External id": 536660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483316.115, "dur": 1.028,
    "args": {
      "External id": 536661,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483316.424, "dur": 0.601,
    "args": {
      "External id": 536662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483321.208, "dur": 6.540,
    "args": {
      "External id": 536663,"Record function id": 0, "Ev Idx": 2134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483322.398, "dur": 2.047,
    "args": {
      "External id": 536664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483322.888, "dur": 1.122,
    "args": {
      "External id": 536665,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483323.252, "dur": 0.654,
    "args": {
      "External id": 536666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483330.902, "dur": 6.128,
    "args": {
      "External id": 536667,"Record function id": 0, "Ev Idx": 2138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483332.067, "dur": 4.589,
    "args": {
      "External id": 536668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483332.521, "dur": 3.717,
    "args": {
      "External id": 536669,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483333.114, "dur": 3.051,
    "args": {
      "External id": 536670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483339.969, "dur": 3.860,
    "args": {
      "External id": 536671,"Record function id": 0, "Ev Idx": 2142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622483341.109, "dur": 2.323,
    "args": {
      "External id": 536672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483341.567, "dur": 1.462,
    "args": {
      "External id": 536673,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622483342.050, "dur": 0.887,
    "args": {
      "External id": 536674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622483347.678, "dur": 15006.321,
    "args": {
      "External id": 536675,"Record function id": 0, "Sequence number": 6319449, "Fwd thread id": 1, "Ev Idx": 2146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622483349.206, "dur": 14995.902,
    "args": {
      "External id": 536676,"Sequence number": 6319449, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2147
    }
  },
  {
    "ph": "f", "id": 179, "pid": 4183438, "tid": 31367, "ts": 679622483349.206,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.16)", "pid": 4183438, "tid": 31367,
    "ts": 679622483381.816, "dur": 44.610,
    "args": {
      "External id": 536677,"Record function id": 0, "Ev Idx": 2148
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.16)", "pid": 4183438, "tid": 31367,
    "ts": 679622483435.018, "dur": 67.876,
    "args": {
      "External id": 536678,"Record function id": 0, "Ev Idx": 2149
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.16)", "pid": 4183438, "tid": 31367,
    "ts": 679622483509.540, "dur": 14827.084,
    "args": {
      "External id": 536679,"Record function id": 0, "Ev Idx": 2150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622483601.219, "dur": 11.182,
    "args": {
      "External id": 536680,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622483622.020, "dur": 5.145,
    "args": {
      "External id": 536681,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622483641.331, "dur": 13877.682,
    "args": {
      "External id": 536682,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622483693.370, "dur": 13814.318,
    "args": {
      "External id": 536683,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622483729.295, "dur": 15.382,
    "args": {
      "External id": 536684,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622483749.002, "dur": 13719.496,
    "args": {
      "External id": 536685,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 2156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622483751.968, "dur": 13715.547,
    "args": {
      "External id": 536686,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 2157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622483757.375, "dur": 5.844,
    "args": {
      "External id": 536687,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622483765.110, "dur": 13697.980,
    "args": {
      "External id": 536688,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 2159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622497622.111, "dur": 12.069,
    "args": {
      "External id": 536689,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 2160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622497626.026, "dur": 7.753,
    "args": {
      "External id": 536690,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622497692.296, "dur": 327.171,
    "args": {
      "External id": 536691,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 2162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622497720.748, "dur": 293.553,
    "args": {
      "External id": 536692,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2163, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622497733.630, "dur": 275.086,
    "args": {
      "External id": 536693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 2164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622498042.014, "dur": 2.311,
    "args": {
      "External id": 536694,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2165, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498120.471, "dur": 7.556,
    "args": {
      "External id": 536695,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498177.284, "dur": 1.371,
    "args": {
      "External id": 536696,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498196.558, "dur": 1.219,
    "args": {
      "External id": 536697,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498210.819, "dur": 1.214,
    "args": {
      "External id": 536698,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498224.937, "dur": 0.959,
    "args": {
      "External id": 536699,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498236.756, "dur": 0.939,
    "args": {
      "External id": 536700,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498247.862, "dur": 0.932,
    "args": {
      "External id": 536701,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498260.089, "dur": 1.845,
    "args": {
      "External id": 536702,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498272.902, "dur": 0.818,
    "args": {
      "External id": 536703,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622498369.831, "dur": 3279.349,
    "args": {
      "External id": 536704,"Record function id": 0, "Ev Idx": 2175
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.15)", "pid": 4183438, "tid": 31367,
    "ts": 679622498389.042, "dur": 1027.387,
    "args": {
      "External id": 536705,"Record function id": 0, "Ev Idx": 2176
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.15)", "pid": 4183438, "tid": 31367,
    "ts": 679622498403.462, "dur": 349.568,
    "args": {
      "External id": 536706,"Record function id": 0, "Ev Idx": 2177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498480.281, "dur": 4.136,
    "args": {
      "External id": 536707,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498487.683, "dur": 1.263,
    "args": {
      "External id": 536708,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498490.769, "dur": 0.932,
    "args": {
      "External id": 536709,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498493.475, "dur": 0.835,
    "args": {
      "External id": 536710,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498495.850, "dur": 0.920,
    "args": {
      "External id": 536711,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498498.548, "dur": 0.887,
    "args": {
      "External id": 536712,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498500.976, "dur": 1.384,
    "args": {
      "External id": 536713,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498504.315, "dur": 2.885,
    "args": {
      "External id": 536714,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498508.930, "dur": 0.936,
    "args": {
      "External id": 536715,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622498511.395, "dur": 0.770,
    "args": {
      "External id": 536716,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622498530.537, "dur": 189.631,
    "args": {
      "External id": 536717,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622498546.928, "dur": 167.241,
    "args": {
      "External id": 536718,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622498563.759, "dur": 13.957,
    "args": {
      "External id": 536719,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622498581.213, "dur": 64.563,
    "args": {
      "External id": 536720,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 2191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622498584.248, "dur": 61.112,
    "args": {
      "External id": 536721,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 2192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498589.146, "dur": 5.837,
    "args": {
      "External id": 536722,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622498596.826, "dur": 47.887,
    "args": {
      "External id": 536723,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 2194
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.14", "pid": 4183438, "tid": 31367,
    "ts": 679622498838.755, "dur": 567.338,
    "args": {
      "External id": 536724,"Record function id": 0, "Ev Idx": 2195
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.14)", "pid": 4183438, "tid": 31367,
    "ts": 679622498856.244, "dur": 537.897,
    "args": {
      "External id": 536725,"Record function id": 0, "Ev Idx": 2196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622498912.251, "dur": 5.571,
    "args": {
      "External id": 536726,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622498932.547, "dur": 29.967,
    "args": {
      "External id": 536727,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498937.462, "dur": 1.928,
    "args": {
      "External id": 536728,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498940.947, "dur": 0.602,
    "args": {
      "External id": 536729,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498944.221, "dur": 2.268,
    "args": {
      "External id": 536730,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498947.346, "dur": 0.673,
    "args": {
      "External id": 536731,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498949.438, "dur": 2.104,
    "args": {
      "External id": 536732,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498952.712, "dur": 0.291,
    "args": {
      "External id": 536733,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498954.037, "dur": 0.338,
    "args": {
      "External id": 536734,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498956.773, "dur": 0.315,
    "args": {
      "External id": 536735,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622498958.332, "dur": 0.484,
    "args": {
      "External id": 536736,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622498973.017, "dur": 30.382,
    "args": {
      "External id": 536737,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622499033.756, "dur": 123.334,
    "args": {
      "External id": 536738,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 2209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622499043.325, "dur": 3.603,
    "args": {
      "External id": 536739,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622499052.649, "dur": 11.919,
    "args": {
      "External id": 536740,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622499056.814, "dur": 7.360,
    "args": {
      "External id": 536741,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 2212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499060.261, "dur": 2.756,
    "args": {
      "External id": 536742,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622499070.670, "dur": 25.300,
    "args": {
      "External id": 536743,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499073.170, "dur": 0.834,
    "args": {
      "External id": 536744,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499076.613, "dur": 0.435,
    "args": {
      "External id": 536745,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499078.274, "dur": 0.752,
    "args": {
      "External id": 536746,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499080.062, "dur": 0.741,
    "args": {
      "External id": 536747,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499083.372, "dur": 0.522,
    "args": {
      "External id": 536748,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499085.044, "dur": 0.614,
    "args": {
      "External id": 536749,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499087.015, "dur": 1.436,
    "args": {
      "External id": 536750,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499089.344, "dur": 1.938,
    "args": {
      "External id": 536751,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622499092.467, "dur": 0.275,
    "args": {
      "External id": 536752,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622499124.009, "dur": 25.205,
    "args": {
      "External id": 536753,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622499206.282, "dur": 121.314,
    "args": {
      "External id": 536754,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 2225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622499231.049, "dur": 92.951,
    "args": {
      "External id": 536755,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2226, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622499240.727, "dur": 79.343,
    "args": {
      "External id": 536756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 2227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622499339.994, "dur": 2.045,
    "args": {
      "External id": 536757,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2228, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622499422.898, "dur": 2198.623,
    "args": {
      "External id": 536758,"Sequence number": 6319448, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2229
    }
  },
  {
    "ph": "f", "id": 180, "pid": 4183438, "tid": 31367, "ts": 679622499422.898,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622499538.624, "dur": 106.025,
    "args": {
      "External id": 536759,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 2230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622499727.879, "dur": 41.646,
    "args": {
      "External id": 536760,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 2231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622499790.332, "dur": 50.516,
    "args": {
      "External id": 536761,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 2232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622499850.391, "dur": 30.311,
    "args": {
      "External id": 536762,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622499886.735, "dur": 33.144,
    "args": {
      "External id": 536763,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622499926.535, "dur": 23.171,
    "args": {
      "External id": 536764,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622499958.422, "dur": 29.783,
    "args": {
      "External id": 536765,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622500011.785, "dur": 21.968,
    "args": {
      "External id": 536766,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 2237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622500052.096, "dur": 25.959,
    "args": {
      "External id": 536767,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622500097.981, "dur": 35.390,
    "args": {
      "External id": 536768,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622500148.404, "dur": 14.190,
    "args": {
      "External id": 536769,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500173.021, "dur": 40.533,
    "args": {
      "External id": 536770,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500217.377, "dur": 30.229,
    "args": {
      "External id": 536771,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622500274.745, "dur": 166.241,
    "args": {
      "External id": 536772,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622500351.042, "dur": 5.456,
    "args": {
      "External id": 536773,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622500358.745, "dur": 2.560,
    "args": {
      "External id": 536774,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622500469.016, "dur": 25.648,
    "args": {
      "External id": 536775,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622500506.050, "dur": 14.886,
    "args": {
      "External id": 536776,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500530.127, "dur": 41.933,
    "args": {
      "External id": 536777,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500577.710, "dur": 29.685,
    "args": {
      "External id": 536778,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500619.565, "dur": 27.684,
    "args": {
      "External id": 536779,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500691.337, "dur": 49.890,
    "args": {
      "External id": 536780,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500756.944, "dur": 38.599,
    "args": {
      "External id": 536781,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622500807.176, "dur": 55.967,
    "args": {
      "External id": 536782,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622500900.834, "dur": 41.788,
    "args": {
      "External id": 536783,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 2254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622500970.138, "dur": 35.713,
    "args": {
      "External id": 536784,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622501235.386, "dur": 96.357,
    "args": {
      "External id": 536785,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622501523.014, "dur": 20.691,
    "args": {
      "External id": 536786,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622501559.191, "dur": 23.986,
    "args": {
      "External id": 536787,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 2258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501703.698, "dur": 18.012,
    "args": {
      "External id": 536788,"Record function id": 0, "Ev Idx": 2259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501707.989, "dur": 12.602,
    "args": {
      "External id": 536789,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501712.836, "dur": 6.609,
    "args": {
      "External id": 536790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501714.725, "dur": 4.503,
    "args": {
      "External id": 536791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501725.889, "dur": 5.344,
    "args": {
      "External id": 536792,"Record function id": 0, "Ev Idx": 2263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501727.598, "dur": 3.225,
    "args": {
      "External id": 536793,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501728.583, "dur": 1.768,
    "args": {
      "External id": 536794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501729.425, "dur": 0.840,
    "args": {
      "External id": 536795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501734.391, "dur": 3.749,
    "args": {
      "External id": 536796,"Record function id": 0, "Ev Idx": 2267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501735.684, "dur": 2.081,
    "args": {
      "External id": 536797,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501736.191, "dur": 1.195,
    "args": {
      "External id": 536798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501736.542, "dur": 0.777,
    "args": {
      "External id": 536799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501741.198, "dur": 3.953,
    "args": {
      "External id": 536800,"Record function id": 0, "Ev Idx": 2271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501742.640, "dur": 2.138,
    "args": {
      "External id": 536801,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501743.116, "dur": 1.260,
    "args": {
      "External id": 536802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501743.626, "dur": 0.685,
    "args": {
      "External id": 536803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501748.244, "dur": 4.333,
    "args": {
      "External id": 536804,"Record function id": 0, "Ev Idx": 2275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501749.617, "dur": 2.581,
    "args": {
      "External id": 536805,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501750.274, "dur": 1.463,
    "args": {
      "External id": 536806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501750.980, "dur": 0.692,
    "args": {
      "External id": 536807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501755.705, "dur": 3.788,
    "args": {
      "External id": 536808,"Record function id": 0, "Ev Idx": 2279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501756.974, "dur": 2.130,
    "args": {
      "External id": 536809,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501757.454, "dur": 1.255,
    "args": {
      "External id": 536810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501758.086, "dur": 0.558,
    "args": {
      "External id": 536811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501762.751, "dur": 5.409,
    "args": {
      "External id": 536812,"Record function id": 0, "Ev Idx": 2283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501763.860, "dur": 3.902,
    "args": {
      "External id": 536813,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501764.399, "dur": 2.969,
    "args": {
      "External id": 536814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501764.958, "dur": 2.339,
    "args": {
      "External id": 536815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501771.175, "dur": 4.001,
    "args": {
      "External id": 536816,"Record function id": 0, "Ev Idx": 2287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501772.296, "dur": 2.488,
    "args": {
      "External id": 536817,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501772.724, "dur": 1.665,
    "args": {
      "External id": 536818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501773.438, "dur": 0.880,
    "args": {
      "External id": 536819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501778.846, "dur": 3.460,
    "args": {
      "External id": 536820,"Record function id": 0, "Ev Idx": 2291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622501779.900, "dur": 2.020,
    "args": {
      "External id": 536821,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501780.360, "dur": 1.181,
    "args": {
      "External id": 536822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622501780.846, "dur": 0.623,
    "args": {
      "External id": 536823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622501786.050, "dur": 14937.914,
    "args": {
      "External id": 536824,"Record function id": 0, "Sequence number": 6319447, "Fwd thread id": 1, "Ev Idx": 2295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622501787.180, "dur": 14927.072,
    "args": {
      "External id": 536825,"Sequence number": 6319447, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2296
    }
  },
  {
    "ph": "f", "id": 181, "pid": 4183438, "tid": 31367, "ts": 679622501787.180,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.15)", "pid": 4183438, "tid": 31367,
    "ts": 679622501821.590, "dur": 43.201,
    "args": {
      "External id": 536826,"Record function id": 0, "Ev Idx": 2297
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.15)", "pid": 4183438, "tid": 31367,
    "ts": 679622501873.223, "dur": 65.063,
    "args": {
      "External id": 536827,"Record function id": 0, "Ev Idx": 2298
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.15)", "pid": 4183438, "tid": 31367,
    "ts": 679622501944.897, "dur": 14760.284,
    "args": {
      "External id": 536828,"Record function id": 0, "Ev Idx": 2299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622502037.681, "dur": 8.229,
    "args": {
      "External id": 536829,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622502056.547, "dur": 4.705,
    "args": {
      "External id": 536830,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622502074.954, "dur": 13893.752,
    "args": {
      "External id": 536831,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622502088.747, "dur": 13869.492,
    "args": {
      "External id": 536832,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622502130.227, "dur": 14.675,
    "args": {
      "External id": 536833,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622502149.356, "dur": 13771.251,
    "args": {
      "External id": 536834,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 2305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622502153.943, "dur": 13765.774,
    "args": {
      "External id": 536835,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 2306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622502158.237, "dur": 5.690,
    "args": {
      "External id": 536836,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622502165.823, "dur": 13745.912,
    "args": {
      "External id": 536837,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 2308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622516072.684, "dur": 11.966,
    "args": {
      "External id": 536838,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 2309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622516075.933, "dur": 8.349,
    "args": {
      "External id": 536839,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622516129.355, "dur": 259.958,
    "args": {
      "External id": 536840,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 2311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622516158.741, "dur": 225.912,
    "args": {
      "External id": 536841,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2312, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622516171.429, "dur": 208.237,
    "args": {
      "External id": 536842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 2313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622516406.853, "dur": 2.170,
    "args": {
      "External id": 536843,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2314, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516464.921, "dur": 8.245,
    "args": {
      "External id": 536844,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516518.170, "dur": 1.267,
    "args": {
      "External id": 536845,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516535.583, "dur": 1.285,
    "args": {
      "External id": 536846,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516547.841, "dur": 1.083,
    "args": {
      "External id": 536847,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516560.884, "dur": 2.808,
    "args": {
      "External id": 536848,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516574.272, "dur": 1.253,
    "args": {
      "External id": 536849,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516585.703, "dur": 0.790,
    "args": {
      "External id": 536850,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516597.354, "dur": 1.874,
    "args": {
      "External id": 536851,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516608.806, "dur": 2.767,
    "args": {
      "External id": 536852,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622516740.518, "dur": 3271.888,
    "args": {
      "External id": 536853,"Record function id": 0, "Ev Idx": 2324
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.14)", "pid": 4183438, "tid": 31367,
    "ts": 679622516759.858, "dur": 1064.575,
    "args": {
      "External id": 536854,"Record function id": 0, "Ev Idx": 2325
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.14)", "pid": 4183438, "tid": 31367,
    "ts": 679622516775.457, "dur": 306.922,
    "args": {
      "External id": 536855,"Record function id": 0, "Ev Idx": 2326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516853.389, "dur": 4.843,
    "args": {
      "External id": 536856,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516861.406, "dur": 0.843,
    "args": {
      "External id": 536857,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516863.968, "dur": 1.190,
    "args": {
      "External id": 536858,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516867.136, "dur": 0.629,
    "args": {
      "External id": 536859,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516869.749, "dur": 1.109,
    "args": {
      "External id": 536860,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516872.472, "dur": 0.800,
    "args": {
      "External id": 536861,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516874.913, "dur": 3.979,
    "args": {
      "External id": 536862,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516880.666, "dur": 1.026,
    "args": {
      "External id": 536863,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516883.447, "dur": 0.647,
    "args": {
      "External id": 536864,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622516885.668, "dur": 0.598,
    "args": {
      "External id": 536865,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622516904.152, "dur": 149.608,
    "args": {
      "External id": 536866,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622516920.160, "dur": 129.222,
    "args": {
      "External id": 536867,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622516936.838, "dur": 13.844,
    "args": {
      "External id": 536868,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622516953.459, "dur": 66.403,
    "args": {
      "External id": 536869,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 2340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622516956.116, "dur": 63.440,
    "args": {
      "External id": 536870,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 2341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622516960.268, "dur": 6.104,
    "args": {
      "External id": 536871,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622516968.581, "dur": 50.156,
    "args": {
      "External id": 536872,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 2343
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.13", "pid": 4183438, "tid": 31367,
    "ts": 679622517191.287, "dur": 625.682,
    "args": {
      "External id": 536873,"Record function id": 0, "Ev Idx": 2344
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.13)", "pid": 4183438, "tid": 31367,
    "ts": 679622517213.045, "dur": 588.738,
    "args": {
      "External id": 536874,"Record function id": 0, "Ev Idx": 2345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622517272.081, "dur": 5.538,
    "args": {
      "External id": 536875,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622517293.678, "dur": 33.097,
    "args": {
      "External id": 536876,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517298.647, "dur": 2.635,
    "args": {
      "External id": 536877,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517303.664, "dur": 2.437,
    "args": {
      "External id": 536878,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517307.347, "dur": 0.524,
    "args": {
      "External id": 536879,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517310.985, "dur": 0.280,
    "args": {
      "External id": 536880,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517312.724, "dur": 0.461,
    "args": {
      "External id": 536881,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517314.806, "dur": 0.551,
    "args": {
      "External id": 536882,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517317.597, "dur": 0.528,
    "args": {
      "External id": 536883,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517319.452, "dur": 0.418,
    "args": {
      "External id": 536884,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517321.575, "dur": 1.022,
    "args": {
      "External id": 536885,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622517338.133, "dur": 32.303,
    "args": {
      "External id": 536886,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622517404.425, "dur": 105.741,
    "args": {
      "External id": 536887,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 2358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622517414.022, "dur": 5.654,
    "args": {
      "External id": 536888,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622517424.941, "dur": 10.449,
    "args": {
      "External id": 536889,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622517429.268, "dur": 5.734,
    "args": {
      "External id": 536890,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 2361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517433.261, "dur": 0.647,
    "args": {
      "External id": 536891,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622517442.154, "dur": 30.609,
    "args": {
      "External id": 536892,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517444.881, "dur": 0.493,
    "args": {
      "External id": 536893,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517447.372, "dur": 1.271,
    "args": {
      "External id": 536894,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517449.968, "dur": 0.518,
    "args": {
      "External id": 536895,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517452.087, "dur": 0.587,
    "args": {
      "External id": 536896,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517455.178, "dur": 0.514,
    "args": {
      "External id": 536897,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517456.999, "dur": 0.541,
    "args": {
      "External id": 536898,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517458.712, "dur": 5.894,
    "args": {
      "External id": 536899,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517466.994, "dur": 0.582,
    "args": {
      "External id": 536900,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622517468.786, "dur": 0.362,
    "args": {
      "External id": 536901,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622517483.783, "dur": 18.670,
    "args": {
      "External id": 536902,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622517555.140, "dur": 166.899,
    "args": {
      "External id": 536903,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 2374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622517583.425, "dur": 134.262,
    "args": {
      "External id": 536904,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2375, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622517592.862, "dur": 119.959,
    "args": {
      "External id": 536905,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 2376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622517739.911, "dur": 2.162,
    "args": {
      "External id": 536906,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2377, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622517832.138, "dur": 2158.470,
    "args": {
      "External id": 536907,"Sequence number": 6319446, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2378
    }
  },
  {
    "ph": "f", "id": 182, "pid": 4183438, "tid": 31367, "ts": 679622517832.138,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622517941.905, "dur": 102.588,
    "args": {
      "External id": 536908,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 2379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622518084.092, "dur": 53.919,
    "args": {
      "External id": 536909,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 2380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622518159.614, "dur": 44.830,
    "args": {
      "External id": 536910,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 2381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622518213.828, "dur": 26.604,
    "args": {
      "External id": 536911,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622518246.954, "dur": 34.799,
    "args": {
      "External id": 536912,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622518290.724, "dur": 21.088,
    "args": {
      "External id": 536913,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622518319.846, "dur": 29.542,
    "args": {
      "External id": 536914,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622518376.100, "dur": 24.803,
    "args": {
      "External id": 536915,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 2386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622518425.518, "dur": 29.849,
    "args": {
      "External id": 536916,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622518474.698, "dur": 19.917,
    "args": {
      "External id": 536917,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622518507.771, "dur": 16.216,
    "args": {
      "External id": 536918,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622518533.619, "dur": 37.134,
    "args": {
      "External id": 536919,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622518574.259, "dur": 33.855,
    "args": {
      "External id": 536920,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622518634.856, "dur": 219.999,
    "args": {
      "External id": 536921,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622518752.939, "dur": 6.840,
    "args": {
      "External id": 536922,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622518762.065, "dur": 5.532,
    "args": {
      "External id": 536923,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622518887.508, "dur": 27.645,
    "args": {
      "External id": 536924,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622518926.912, "dur": 14.456,
    "args": {
      "External id": 536925,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622518949.414, "dur": 52.180,
    "args": {
      "External id": 536926,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622519007.399, "dur": 35.611,
    "args": {
      "External id": 536927,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622519051.840, "dur": 35.082,
    "args": {
      "External id": 536928,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622519112.901, "dur": 49.748,
    "args": {
      "External id": 536929,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622519178.274, "dur": 39.763,
    "args": {
      "External id": 536930,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622519232.934, "dur": 61.305,
    "args": {
      "External id": 536931,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622519339.283, "dur": 37.291,
    "args": {
      "External id": 536932,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 2403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622519418.181, "dur": 191.209,
    "args": {
      "External id": 536933,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622519629.592, "dur": 221.990,
    "args": {
      "External id": 536934,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622519880.741, "dur": 22.383,
    "args": {
      "External id": 536935,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622519932.133, "dur": 23.209,
    "args": {
      "External id": 536936,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 2407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520036.611, "dur": 16.433,
    "args": {
      "External id": 536937,"Record function id": 0, "Ev Idx": 2408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520040.084, "dur": 12.003,
    "args": {
      "External id": 536938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520045.131, "dur": 6.222,
    "args": {
      "External id": 536939,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520046.953, "dur": 4.317,
    "args": {
      "External id": 536940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520056.975, "dur": 5.302,
    "args": {
      "External id": 536941,"Record function id": 0, "Ev Idx": 2412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520058.531, "dur": 3.340,
    "args": {
      "External id": 536942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520059.426, "dur": 1.993,
    "args": {
      "External id": 536943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520060.194, "dur": 1.133,
    "args": {
      "External id": 536944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520065.722, "dur": 4.820,
    "args": {
      "External id": 536945,"Record function id": 0, "Ev Idx": 2416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520067.410, "dur": 2.713,
    "args": {
      "External id": 536946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520068.093, "dur": 1.619,
    "args": {
      "External id": 536947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520068.710, "dur": 0.923,
    "args": {
      "External id": 536948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520073.644, "dur": 5.911,
    "args": {
      "External id": 536949,"Record function id": 0, "Ev Idx": 2420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520074.752, "dur": 4.412,
    "args": {
      "External id": 536950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520075.527, "dur": 3.230,
    "args": {
      "External id": 536951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520076.015, "dur": 2.675,
    "args": {
      "External id": 536952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520082.954, "dur": 3.783,
    "args": {
      "External id": 536953,"Record function id": 0, "Ev Idx": 2424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520084.073, "dur": 2.293,
    "args": {
      "External id": 536954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520084.570, "dur": 1.408,
    "args": {
      "External id": 536955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520085.127, "dur": 0.777,
    "args": {
      "External id": 536956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520090.077, "dur": 3.961,
    "args": {
      "External id": 536957,"Record function id": 0, "Ev Idx": 2428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520091.488, "dur": 2.182,
    "args": {
      "External id": 536958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520091.949, "dur": 1.319,
    "args": {
      "External id": 536959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520092.619, "dur": 0.585,
    "args": {
      "External id": 536960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520097.454, "dur": 20.599,
    "args": {
      "External id": 536961,"Record function id": 0, "Ev Idx": 2432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520112.600, "dur": 4.513,
    "args": {
      "External id": 536962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520113.890, "dur": 2.398,
    "args": {
      "External id": 536963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520114.706, "dur": 1.327,
    "args": {
      "External id": 536964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520122.889, "dur": 4.839,
    "args": {
      "External id": 536965,"Record function id": 0, "Ev Idx": 2436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520124.293, "dur": 3.054,
    "args": {
      "External id": 536966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520125.177, "dur": 1.780,
    "args": {
      "External id": 536967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520126.039, "dur": 0.853,
    "args": {
      "External id": 536968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520131.088, "dur": 4.191,
    "args": {
      "External id": 536969,"Record function id": 0, "Ev Idx": 2440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622520132.386, "dur": 2.496,
    "args": {
      "External id": 536970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520132.893, "dur": 1.602,
    "args": {
      "External id": 536971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622520133.559, "dur": 0.870,
    "args": {
      "External id": 536972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622520139.657, "dur": 16656.840,
    "args": {
      "External id": 536973,"Record function id": 0, "Sequence number": 6319445, "Fwd thread id": 1, "Ev Idx": 2444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622520140.968, "dur": 16642.987,
    "args": {
      "External id": 536974,"Sequence number": 6319445, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2445
    }
  },
  {
    "ph": "f", "id": 183, "pid": 4183438, "tid": 31367, "ts": 679622520140.968,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.14)", "pid": 4183438, "tid": 31367,
    "ts": 679622520173.392, "dur": 42.169,
    "args": {
      "External id": 536975,"Record function id": 0, "Ev Idx": 2446
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.14)", "pid": 4183438, "tid": 31367,
    "ts": 679622520223.826, "dur": 65.255,
    "args": {
      "External id": 536976,"Record function id": 0, "Ev Idx": 2447
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.14)", "pid": 4183438, "tid": 31367,
    "ts": 679622520296.085, "dur": 16479.309,
    "args": {
      "External id": 536977,"Record function id": 0, "Ev Idx": 2448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622520389.862, "dur": 8.280,
    "args": {
      "External id": 536978,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622520408.599, "dur": 4.930,
    "args": {
      "External id": 536979,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622520427.990, "dur": 15516.923,
    "args": {
      "External id": 536980,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622520441.745, "dur": 15490.377,
    "args": {
      "External id": 536981,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622520466.060, "dur": 17.085,
    "args": {
      "External id": 536982,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622520487.132, "dur": 15392.795,
    "args": {
      "External id": 536983,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 2454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622520489.802, "dur": 15389.083,
    "args": {
      "External id": 536984,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 2455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622520494.486, "dur": 4.939,
    "args": {
      "External id": 536985,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622520501.336, "dur": 15372.027,
    "args": {
      "External id": 536986,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 2457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622536053.417, "dur": 12.807,
    "args": {
      "External id": 536987,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 2458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622536057.313, "dur": 8.584,
    "args": {
      "External id": 536988,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622536114.141, "dur": 321.693,
    "args": {
      "External id": 536989,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 2460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622536147.084, "dur": 284.141,
    "args": {
      "External id": 536990,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2461, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622536160.789, "dur": 265.536,
    "args": {
      "External id": 536991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 2462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622536453.717, "dur": 2.174,
    "args": {
      "External id": 536992,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2463, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536514.162, "dur": 7.140,
    "args": {
      "External id": 536993,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536570.494, "dur": 1.510,
    "args": {
      "External id": 536994,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536588.112, "dur": 1.303,
    "args": {
      "External id": 536995,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536606.482, "dur": 1.001,
    "args": {
      "External id": 536996,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536617.591, "dur": 0.748,
    "args": {
      "External id": 536997,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536628.548, "dur": 0.720,
    "args": {
      "External id": 536998,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536640.760, "dur": 0.938,
    "args": {
      "External id": 536999,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536690.316, "dur": 2.931,
    "args": {
      "External id": 537000,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622536708.341, "dur": 1.033,
    "args": {
      "External id": 537001,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622536812.942, "dur": 3198.349,
    "args": {
      "External id": 537002,"Record function id": 0, "Ev Idx": 2473
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.13)", "pid": 4183438, "tid": 31367,
    "ts": 679622536831.911, "dur": 1040.240,
    "args": {
      "External id": 537003,"Record function id": 0, "Ev Idx": 2474
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.13)", "pid": 4183438, "tid": 31367,
    "ts": 679622536847.922, "dur": 343.056,
    "args": {
      "External id": 537004,"Record function id": 0, "Ev Idx": 2475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536928.378, "dur": 4.333,
    "args": {
      "External id": 537005,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536936.528, "dur": 1.090,
    "args": {
      "External id": 537006,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536939.638, "dur": 1.243,
    "args": {
      "External id": 537007,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536942.966, "dur": 2.234,
    "args": {
      "External id": 537008,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536946.874, "dur": 0.797,
    "args": {
      "External id": 537009,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536949.439, "dur": 0.850,
    "args": {
      "External id": 537010,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536952.069, "dur": 1.632,
    "args": {
      "External id": 537011,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536955.380, "dur": 0.791,
    "args": {
      "External id": 537012,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536957.861, "dur": 0.880,
    "args": {
      "External id": 537013,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622536960.479, "dur": 0.858,
    "args": {
      "External id": 537014,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622536979.408, "dur": 180.651,
    "args": {
      "External id": 537015,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622537001.220, "dur": 153.412,
    "args": {
      "External id": 537016,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622537015.763, "dur": 13.933,
    "args": {
      "External id": 537017,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622537032.329, "dur": 89.448,
    "args": {
      "External id": 537018,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 2489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622537034.914, "dur": 86.546,
    "args": {
      "External id": 537019,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 2490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537038.725, "dur": 7.537,
    "args": {
      "External id": 537020,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622537048.048, "dur": 72.385,
    "args": {
      "External id": 537021,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 2492
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.12", "pid": 4183438, "tid": 31367,
    "ts": 679622537270.963, "dur": 593.599,
    "args": {
      "External id": 537022,"Record function id": 0, "Ev Idx": 2493
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.12)", "pid": 4183438, "tid": 31367,
    "ts": 679622537288.741, "dur": 560.683,
    "args": {
      "External id": 537023,"Record function id": 0, "Ev Idx": 2494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622537344.586, "dur": 5.738,
    "args": {
      "External id": 537024,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622537364.976, "dur": 31.535,
    "args": {
      "External id": 537025,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537369.808, "dur": 1.877,
    "args": {
      "External id": 537026,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537374.347, "dur": 0.691,
    "args": {
      "External id": 537027,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537376.820, "dur": 0.364,
    "args": {
      "External id": 537028,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537379.110, "dur": 0.257,
    "args": {
      "External id": 537029,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537381.067, "dur": 0.591,
    "args": {
      "External id": 537030,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537382.829, "dur": 1.259,
    "args": {
      "External id": 537031,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537385.660, "dur": 2.002,
    "args": {
      "External id": 537032,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537389.032, "dur": 0.283,
    "args": {
      "External id": 537033,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537392.151, "dur": 0.233,
    "args": {
      "External id": 537034,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622537406.468, "dur": 32.187,
    "args": {
      "External id": 537035,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622537468.975, "dur": 99.406,
    "args": {
      "External id": 537036,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 2507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622537478.108, "dur": 3.490,
    "args": {
      "External id": 537037,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622537486.914, "dur": 9.946,
    "args": {
      "External id": 537038,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622537490.946, "dur": 5.497,
    "args": {
      "External id": 537039,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 2510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537494.494, "dur": 0.751,
    "args": {
      "External id": 537040,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622537504.003, "dur": 26.602,
    "args": {
      "External id": 537041,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537506.137, "dur": 0.637,
    "args": {
      "External id": 537042,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537508.382, "dur": 0.393,
    "args": {
      "External id": 537043,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537511.325, "dur": 0.572,
    "args": {
      "External id": 537044,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537513.326, "dur": 1.997,
    "args": {
      "External id": 537045,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537516.857, "dur": 0.321,
    "args": {
      "External id": 537046,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537519.515, "dur": 0.575,
    "args": {
      "External id": 537047,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537521.497, "dur": 0.472,
    "args": {
      "External id": 537048,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537523.406, "dur": 1.235,
    "args": {
      "External id": 537049,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622537526.485, "dur": 0.427,
    "args": {
      "External id": 537050,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622537540.532, "dur": 20.929,
    "args": {
      "External id": 537051,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622537609.963, "dur": 164.584,
    "args": {
      "External id": 537052,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 2523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622537636.277, "dur": 134.040,
    "args": {
      "External id": 537053,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2524, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622537645.079, "dur": 120.846,
    "args": {
      "External id": 537054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 2525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622537792.871, "dur": 2.139,
    "args": {
      "External id": 537055,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2526, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622537879.281, "dur": 2109.822,
    "args": {
      "External id": 537056,"Sequence number": 6319444, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2527
    }
  },
  {
    "ph": "f", "id": 184, "pid": 4183438, "tid": 31367, "ts": 679622537879.281,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622537995.662, "dur": 123.629,
    "args": {
      "External id": 537057,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 2528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622538167.075, "dur": 38.204,
    "args": {
      "External id": 537058,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 2529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622538225.527, "dur": 47.517,
    "args": {
      "External id": 537059,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 2530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622538282.942, "dur": 25.058,
    "args": {
      "External id": 537060,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622538315.845, "dur": 35.793,
    "args": {
      "External id": 537061,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622538358.618, "dur": 19.870,
    "args": {
      "External id": 537062,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622538385.421, "dur": 31.471,
    "args": {
      "External id": 537063,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622538440.448, "dur": 21.967,
    "args": {
      "External id": 537064,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 2535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622538481.593, "dur": 27.523,
    "args": {
      "External id": 537065,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622538528.886, "dur": 17.994,
    "args": {
      "External id": 537066,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622538559.544, "dur": 14.105,
    "args": {
      "External id": 537067,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622538581.863, "dur": 35.817,
    "args": {
      "External id": 537068,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622538620.832, "dur": 29.960,
    "args": {
      "External id": 537069,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622538721.660, "dur": 172.161,
    "args": {
      "External id": 537070,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622538800.772, "dur": 6.447,
    "args": {
      "External id": 537071,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622538809.586, "dur": 3.728,
    "args": {
      "External id": 537072,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622538931.803, "dur": 23.947,
    "args": {
      "External id": 537073,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622538968.335, "dur": 13.896,
    "args": {
      "External id": 537074,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622538991.155, "dur": 50.614,
    "args": {
      "External id": 537075,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622539048.012, "dur": 37.947,
    "args": {
      "External id": 537076,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622539113.850, "dur": 48.419,
    "args": {
      "External id": 537077,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622539172.399, "dur": 39.208,
    "args": {
      "External id": 537078,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622539222.149, "dur": 57.821,
    "args": {
      "External id": 537079,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622539298.249, "dur": 46.096,
    "args": {
      "External id": 537080,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622539377.085, "dur": 201.208,
    "args": {
      "External id": 537081,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 2552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622539800.751, "dur": 37.003,
    "args": {
      "External id": 537082,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622539861.783, "dur": 22.779,
    "args": {
      "External id": 537083,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622539910.094, "dur": 15.378,
    "args": {
      "External id": 537084,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622539938.224, "dur": 17.350,
    "args": {
      "External id": 537085,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 2556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540035.051, "dur": 33.572,
    "args": {
      "External id": 537086,"Record function id": 0, "Ev Idx": 2557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540038.568, "dur": 29.000,
    "args": {
      "External id": 537087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540043.311, "dur": 23.216,
    "args": {
      "External id": 537088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540062.018, "dur": 4.413,
    "args": {
      "External id": 537089,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540073.223, "dur": 3.992,
    "args": {
      "External id": 537090,"Record function id": 0, "Ev Idx": 2561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540074.632, "dur": 2.146,
    "args": {
      "External id": 537091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540075.162, "dur": 1.106,
    "args": {
      "External id": 537092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540075.501, "dur": 0.677,
    "args": {
      "External id": 537093,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540081.361, "dur": 5.333,
    "args": {
      "External id": 537094,"Record function id": 0, "Ev Idx": 2565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540082.378, "dur": 3.903,
    "args": {
      "External id": 537095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540082.929, "dur": 2.934,
    "args": {
      "External id": 537096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540083.328, "dur": 2.467,
    "args": {
      "External id": 537097,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540090.454, "dur": 3.411,
    "args": {
      "External id": 537098,"Record function id": 0, "Ev Idx": 2569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540091.369, "dur": 2.093,
    "args": {
      "External id": 537099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540091.917, "dur": 1.157,
    "args": {
      "External id": 537100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540092.242, "dur": 0.764,
    "args": {
      "External id": 537101,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540097.367, "dur": 19.873,
    "args": {
      "External id": 537102,"Record function id": 0, "Ev Idx": 2573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540098.270, "dur": 17.935,
    "args": {
      "External id": 537103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540113.092, "dur": 2.248,
    "args": {
      "External id": 537104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540113.771, "dur": 1.317,
    "args": {
      "External id": 537105,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540122.917, "dur": 4.827,
    "args": {
      "External id": 537106,"Record function id": 0, "Ev Idx": 2577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540124.677, "dur": 2.634,
    "args": {
      "External id": 537107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540125.230, "dur": 1.646,
    "args": {
      "External id": 537108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540126.146, "dur": 0.611,
    "args": {
      "External id": 537109,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540131.515, "dur": 3.973,
    "args": {
      "External id": 537110,"Record function id": 0, "Ev Idx": 2581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540132.736, "dur": 2.303,
    "args": {
      "External id": 537111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540133.221, "dur": 1.392,
    "args": {
      "External id": 537112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540133.853, "dur": 0.695,
    "args": {
      "External id": 537113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540138.953, "dur": 3.761,
    "args": {
      "External id": 537114,"Record function id": 0, "Ev Idx": 2585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540140.131, "dur": 2.191,
    "args": {
      "External id": 537115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540140.606, "dur": 1.307,
    "args": {
      "External id": 537116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540141.228, "dur": 0.620,
    "args": {
      "External id": 537117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540146.425, "dur": 3.500,
    "args": {
      "External id": 537118,"Record function id": 0, "Ev Idx": 2589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622540147.399, "dur": 2.101,
    "args": {
      "External id": 537119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540147.838, "dur": 1.261,
    "args": {
      "External id": 537120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622540148.359, "dur": 0.674,
    "args": {
      "External id": 537121,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622540154.418, "dur": 14897.889,
    "args": {
      "External id": 537122,"Record function id": 0, "Sequence number": 6319443, "Fwd thread id": 1, "Ev Idx": 2593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622540155.834, "dur": 14888.483,
    "args": {
      "External id": 537123,"Sequence number": 6319443, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2594
    }
  },
  {
    "ph": "f", "id": 185, "pid": 4183438, "tid": 31367, "ts": 679622540155.834,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.13)", "pid": 4183438, "tid": 31367,
    "ts": 679622540187.772, "dur": 43.045,
    "args": {
      "External id": 537124,"Record function id": 0, "Ev Idx": 2595
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.13)", "pid": 4183438, "tid": 31367,
    "ts": 679622540238.857, "dur": 64.723,
    "args": {
      "External id": 537125,"Record function id": 0, "Ev Idx": 2596
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.13)", "pid": 4183438, "tid": 31367,
    "ts": 679622540309.763, "dur": 14726.160,
    "args": {
      "External id": 537126,"Record function id": 0, "Ev Idx": 2597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622540402.294, "dur": 7.922,
    "args": {
      "External id": 537127,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622540419.615, "dur": 6.412,
    "args": {
      "External id": 537128,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622540442.197, "dur": 13844.891,
    "args": {
      "External id": 537129,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622540456.363, "dur": 13818.862,
    "args": {
      "External id": 537130,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622540483.539, "dur": 14.005,
    "args": {
      "External id": 537131,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622540501.785, "dur": 13727.724,
    "args": {
      "External id": 537132,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 2603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622540504.112, "dur": 13724.385,
    "args": {
      "External id": 537133,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 2604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622540508.454, "dur": 6.106,
    "args": {
      "External id": 537134,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622540516.344, "dur": 13706.833,
    "args": {
      "External id": 537135,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 2606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622554396.978, "dur": 12.326,
    "args": {
      "External id": 537136,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 2607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622554400.459, "dur": 8.518,
    "args": {
      "External id": 537137,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622554439.783, "dur": 308.319,
    "args": {
      "External id": 537138,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 2609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622554467.921, "dur": 275.063,
    "args": {
      "External id": 537139,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2610, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622554479.950, "dur": 256.327,
    "args": {
      "External id": 537140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 2611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622554771.500, "dur": 2.167,
    "args": {
      "External id": 537141,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2612, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554831.907, "dur": 6.940,
    "args": {
      "External id": 537142,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554883.017, "dur": 1.629,
    "args": {
      "External id": 537143,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554900.447, "dur": 3.120,
    "args": {
      "External id": 537144,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554915.337, "dur": 0.914,
    "args": {
      "External id": 537145,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554927.118, "dur": 0.809,
    "args": {
      "External id": 537146,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554938.042, "dur": 0.923,
    "args": {
      "External id": 537147,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554949.777, "dur": 2.738,
    "args": {
      "External id": 537148,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554963.425, "dur": 1.726,
    "args": {
      "External id": 537149,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622554976.236, "dur": 0.769,
    "args": {
      "External id": 537150,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622555066.879, "dur": 3196.551,
    "args": {
      "External id": 537151,"Record function id": 0, "Ev Idx": 2622
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.12)", "pid": 4183438, "tid": 31367,
    "ts": 679622555085.557, "dur": 1052.301,
    "args": {
      "External id": 537152,"Record function id": 0, "Ev Idx": 2623
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.12)", "pid": 4183438, "tid": 31367,
    "ts": 679622555115.783, "dur": 313.194,
    "args": {
      "External id": 537153,"Record function id": 0, "Ev Idx": 2624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555198.355, "dur": 4.801,
    "args": {
      "External id": 537154,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555206.469, "dur": 0.969,
    "args": {
      "External id": 537155,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555209.412, "dur": 2.885,
    "args": {
      "External id": 537156,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555214.263, "dur": 0.662,
    "args": {
      "External id": 537157,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555216.892, "dur": 0.991,
    "args": {
      "External id": 537158,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555219.343, "dur": 0.652,
    "args": {
      "External id": 537159,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555221.835, "dur": 1.708,
    "args": {
      "External id": 537160,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555225.184, "dur": 0.863,
    "args": {
      "External id": 537161,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555228.067, "dur": 0.772,
    "args": {
      "External id": 537162,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622555230.438, "dur": 0.769,
    "args": {
      "External id": 537163,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622555250.077, "dur": 149.383,
    "args": {
      "External id": 537164,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622555267.076, "dur": 127.890,
    "args": {
      "External id": 537165,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622555282.110, "dur": 14.149,
    "args": {
      "External id": 537166,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622555298.998, "dur": 66.394,
    "args": {
      "External id": 537167,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 2638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622555301.601, "dur": 63.451,
    "args": {
      "External id": 537168,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 2639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555306.656, "dur": 9.400,
    "args": {
      "External id": 537169,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622555317.958, "dur": 46.311,
    "args": {
      "External id": 537170,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 2641
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.11", "pid": 4183438, "tid": 31367,
    "ts": 679622555504.118, "dur": 625.224,
    "args": {
      "External id": 537171,"Record function id": 0, "Ev Idx": 2642
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.11)", "pid": 4183438, "tid": 31367,
    "ts": 679622555519.073, "dur": 575.046,
    "args": {
      "External id": 537172,"Record function id": 0, "Ev Idx": 2643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622555572.109, "dur": 5.389,
    "args": {
      "External id": 537173,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622555591.869, "dur": 29.683,
    "args": {
      "External id": 537174,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555596.602, "dur": 1.549,
    "args": {
      "External id": 537175,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555600.500, "dur": 0.460,
    "args": {
      "External id": 537176,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555603.077, "dur": 0.489,
    "args": {
      "External id": 537177,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555604.947, "dur": 0.595,
    "args": {
      "External id": 537178,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555606.704, "dur": 1.507,
    "args": {
      "External id": 537179,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555609.967, "dur": 2.526,
    "args": {
      "External id": 537180,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555613.557, "dur": 0.420,
    "args": {
      "External id": 537181,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555616.279, "dur": 0.329,
    "args": {
      "External id": 537182,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555617.346, "dur": 0.460,
    "args": {
      "External id": 537183,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622555631.206, "dur": 72.130,
    "args": {
      "External id": 537184,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622555741.793, "dur": 110.892,
    "args": {
      "External id": 537185,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 2656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622555752.414, "dur": 5.280,
    "args": {
      "External id": 537186,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622555763.169, "dur": 14.103,
    "args": {
      "External id": 537187,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622555767.898, "dur": 8.982,
    "args": {
      "External id": 537188,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 2659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555774.418, "dur": 0.812,
    "args": {
      "External id": 537189,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622555784.158, "dur": 24.280,
    "args": {
      "External id": 537190,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555786.105, "dur": 0.603,
    "args": {
      "External id": 537191,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555789.541, "dur": 0.273,
    "args": {
      "External id": 537192,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555791.078, "dur": 2.116,
    "args": {
      "External id": 537193,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555794.297, "dur": 0.390,
    "args": {
      "External id": 537194,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555797.205, "dur": 0.569,
    "args": {
      "External id": 537195,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555799.071, "dur": 0.366,
    "args": {
      "External id": 537196,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555800.983, "dur": 1.166,
    "args": {
      "External id": 537197,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555803.264, "dur": 0.376,
    "args": {
      "External id": 537198,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622555804.858, "dur": 0.306,
    "args": {
      "External id": 537199,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622555820.618, "dur": 24.592,
    "args": {
      "External id": 537200,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622555902.128, "dur": 120.674,
    "args": {
      "External id": 537201,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 2672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622555928.254, "dur": 91.119,
    "args": {
      "External id": 537202,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2673, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622555938.092, "dur": 77.221,
    "args": {
      "External id": 537203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 2674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622556038.319, "dur": 1.845,
    "args": {
      "External id": 537204,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2675, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622556145.281, "dur": 2096.495,
    "args": {
      "External id": 537205,"Sequence number": 6319442, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2676
    }
  },
  {
    "ph": "f", "id": 186, "pid": 4183438, "tid": 31367, "ts": 679622556145.281,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622556259.931, "dur": 110.161,
    "args": {
      "External id": 537206,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 2677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622556411.915, "dur": 40.086,
    "args": {
      "External id": 537207,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 2678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622556468.491, "dur": 45.239,
    "args": {
      "External id": 537208,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 2679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622556524.866, "dur": 26.465,
    "args": {
      "External id": 537209,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622556557.470, "dur": 33.077,
    "args": {
      "External id": 537210,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622556598.192, "dur": 21.949,
    "args": {
      "External id": 537211,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622556627.443, "dur": 69.105,
    "args": {
      "External id": 537212,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622556727.819, "dur": 26.166,
    "args": {
      "External id": 537213,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 2684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622556772.609, "dur": 30.627,
    "args": {
      "External id": 537214,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622556823.251, "dur": 19.165,
    "args": {
      "External id": 537215,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622556858.031, "dur": 15.480,
    "args": {
      "External id": 537216,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622556881.962, "dur": 41.441,
    "args": {
      "External id": 537217,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622556926.953, "dur": 31.218,
    "args": {
      "External id": 537218,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622556987.603, "dur": 183.861,
    "args": {
      "External id": 537219,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622557065.180, "dur": 6.250,
    "args": {
      "External id": 537220,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622557073.389, "dur": 2.873,
    "args": {
      "External id": 537221,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622557203.759, "dur": 23.051,
    "args": {
      "External id": 537222,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622557237.959, "dur": 13.242,
    "args": {
      "External id": 537223,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622557259.934, "dur": 46.386,
    "args": {
      "External id": 537224,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622557311.845, "dur": 30.375,
    "args": {
      "External id": 537225,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622557351.233, "dur": 26.718,
    "args": {
      "External id": 537226,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622557384.063, "dur": 38.035,
    "args": {
      "External id": 537227,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622557430.926, "dur": 41.077,
    "args": {
      "External id": 537228,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622557481.922, "dur": 39.610,
    "args": {
      "External id": 537229,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622557550.849, "dur": 52.264,
    "args": {
      "External id": 537230,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 2701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622557639.720, "dur": 69.463,
    "args": {
      "External id": 537231,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622557735.683, "dur": 192.674,
    "args": {
      "External id": 537232,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622557952.302, "dur": 202.202,
    "args": {
      "External id": 537233,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622558177.215, "dur": 26.176,
    "args": {
      "External id": 537234,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 2705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558287.589, "dur": 17.006,
    "args": {
      "External id": 537235,"Record function id": 0, "Ev Idx": 2706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558290.728, "dur": 12.918,
    "args": {
      "External id": 537236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558295.410, "dur": 7.456,
    "args": {
      "External id": 537237,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558296.960, "dur": 5.801,
    "args": {
      "External id": 537238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558308.476, "dur": 6.495,
    "args": {
      "External id": 537239,"Record function id": 0, "Ev Idx": 2710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558310.018, "dur": 4.517,
    "args": {
      "External id": 537240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558310.835, "dur": 3.227,
    "args": {
      "External id": 537241,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558311.473, "dur": 2.409,
    "args": {
      "External id": 537242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558318.221, "dur": 4.703,
    "args": {
      "External id": 537243,"Record function id": 0, "Ev Idx": 2714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558320.184, "dur": 2.356,
    "args": {
      "External id": 537244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558320.860, "dur": 1.253,
    "args": {
      "External id": 537245,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558321.365, "dur": 0.670,
    "args": {
      "External id": 537246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558326.085, "dur": 3.761,
    "args": {
      "External id": 537247,"Record function id": 0, "Ev Idx": 2718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558327.309, "dur": 2.138,
    "args": {
      "External id": 537248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558327.804, "dur": 1.240,
    "args": {
      "External id": 537249,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558328.261, "dur": 0.658,
    "args": {
      "External id": 537250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558332.897, "dur": 4.051,
    "args": {
      "External id": 537251,"Record function id": 0, "Ev Idx": 2722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558334.263, "dur": 2.306,
    "args": {
      "External id": 537252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558334.815, "dur": 1.356,
    "args": {
      "External id": 537253,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558335.504, "dur": 0.567,
    "args": {
      "External id": 537254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558340.069, "dur": 3.731,
    "args": {
      "External id": 537255,"Record function id": 0, "Ev Idx": 2726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558341.236, "dur": 2.171,
    "args": {
      "External id": 537256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558341.663, "dur": 1.341,
    "args": {
      "External id": 537257,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558342.215, "dur": 0.697,
    "args": {
      "External id": 537258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558347.076, "dur": 4.056,
    "args": {
      "External id": 537259,"Record function id": 0, "Ev Idx": 2730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558348.460, "dur": 2.268,
    "args": {
      "External id": 537260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558348.917, "dur": 1.404,
    "args": {
      "External id": 537261,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558349.715, "dur": 0.503,
    "args": {
      "External id": 537262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558354.190, "dur": 3.601,
    "args": {
      "External id": 537263,"Record function id": 0, "Ev Idx": 2734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558355.263, "dur": 2.158,
    "args": {
      "External id": 537264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558355.768, "dur": 1.252,
    "args": {
      "External id": 537265,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558356.346, "dur": 0.582,
    "args": {
      "External id": 537266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558360.727, "dur": 6.205,
    "args": {
      "External id": 537267,"Record function id": 0, "Ev Idx": 2738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622558361.857, "dur": 4.671,
    "args": {
      "External id": 537268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558362.414, "dur": 3.710,
    "args": {
      "External id": 537269,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622558363.077, "dur": 2.977,
    "args": {
      "External id": 537270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622558370.575, "dur": 14969.883,
    "args": {
      "External id": 537271,"Record function id": 0, "Sequence number": 6319441, "Fwd thread id": 1, "Ev Idx": 2742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622558371.768, "dur": 14959.808,
    "args": {
      "External id": 537272,"Sequence number": 6319441, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2743
    }
  },
  {
    "ph": "f", "id": 187, "pid": 4183438, "tid": 31367, "ts": 679622558371.768,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.12)", "pid": 4183438, "tid": 31367,
    "ts": 679622558404.058, "dur": 39.965,
    "args": {
      "External id": 537273,"Record function id": 0, "Ev Idx": 2744
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.12)", "pid": 4183438, "tid": 31367,
    "ts": 679622558452.256, "dur": 63.470,
    "args": {
      "External id": 537274,"Record function id": 0, "Ev Idx": 2745
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.12)", "pid": 4183438, "tid": 31367,
    "ts": 679622558523.211, "dur": 14800.738,
    "args": {
      "External id": 537275,"Record function id": 0, "Ev Idx": 2746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622558612.910, "dur": 8.569,
    "args": {
      "External id": 537276,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622558630.690, "dur": 5.043,
    "args": {
      "External id": 537277,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622558648.930, "dur": 13957.183,
    "args": {
      "External id": 537278,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622558699.076, "dur": 13896.621,
    "args": {
      "External id": 537279,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622558730.988, "dur": 15.023,
    "args": {
      "External id": 537280,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622558750.112, "dur": 13802.564,
    "args": {
      "External id": 537281,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 2752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622558752.484, "dur": 13799.303,
    "args": {
      "External id": 537282,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 2753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622558756.878, "dur": 7.507,
    "args": {
      "External id": 537283,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622558765.925, "dur": 13781.092,
    "args": {
      "External id": 537284,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 2755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622572741.250, "dur": 11.532,
    "args": {
      "External id": 537285,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 2756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622572744.510, "dur": 7.666,
    "args": {
      "External id": 537286,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622572787.558, "dur": 251.789,
    "args": {
      "External id": 537287,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 2758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622572814.835, "dur": 220.211,
    "args": {
      "External id": 537288,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2759, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622572827.632, "dur": 202.767,
    "args": {
      "External id": 537289,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 2760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622573053.684, "dur": 2.312,
    "args": {
      "External id": 537290,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2761, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573125.401, "dur": 6.918,
    "args": {
      "External id": 537291,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573177.943, "dur": 3.189,
    "args": {
      "External id": 537292,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573197.275, "dur": 1.113,
    "args": {
      "External id": 537293,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573209.152, "dur": 1.027,
    "args": {
      "External id": 537294,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573220.214, "dur": 0.696,
    "args": {
      "External id": 537295,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573229.474, "dur": 2.151,
    "args": {
      "External id": 537296,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573242.029, "dur": 0.753,
    "args": {
      "External id": 537297,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573253.050, "dur": 1.998,
    "args": {
      "External id": 537298,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573264.306, "dur": 0.671,
    "args": {
      "External id": 537299,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622573356.335, "dur": 3267.002,
    "args": {
      "External id": 537300,"Record function id": 0, "Ev Idx": 2771
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.11)", "pid": 4183438, "tid": 31367,
    "ts": 679622573374.852, "dur": 1025.537,
    "args": {
      "External id": 537301,"Record function id": 0, "Ev Idx": 2772
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.11)", "pid": 4183438, "tid": 31367,
    "ts": 679622573388.725, "dur": 347.370,
    "args": {
      "External id": 537302,"Record function id": 0, "Ev Idx": 2773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573464.308, "dur": 5.741,
    "args": {
      "External id": 537303,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573473.103, "dur": 0.896,
    "args": {
      "External id": 537304,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573475.854, "dur": 0.801,
    "args": {
      "External id": 537305,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573478.236, "dur": 0.975,
    "args": {
      "External id": 537306,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573480.897, "dur": 0.638,
    "args": {
      "External id": 537307,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573484.382, "dur": 0.785,
    "args": {
      "External id": 537308,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573486.728, "dur": 1.341,
    "args": {
      "External id": 537309,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573489.552, "dur": 0.993,
    "args": {
      "External id": 537310,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573491.877, "dur": 2.425,
    "args": {
      "External id": 537311,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622573495.691, "dur": 0.654,
    "args": {
      "External id": 537312,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622573516.652, "dur": 186.742,
    "args": {
      "External id": 537313,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622573532.644, "dur": 164.827,
    "args": {
      "External id": 537314,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622573545.206, "dur": 17.276,
    "args": {
      "External id": 537315,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622573565.303, "dur": 63.990,
    "args": {
      "External id": 537316,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 2787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622573567.751, "dur": 61.222,
    "args": {
      "External id": 537317,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 2788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573572.873, "dur": 5.149,
    "args": {
      "External id": 537318,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622573579.750, "dur": 48.352,
    "args": {
      "External id": 537319,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 2790
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.10", "pid": 4183438, "tid": 31367,
    "ts": 679622573816.590, "dur": 575.429,
    "args": {
      "External id": 537320,"Record function id": 0, "Ev Idx": 2791
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.10)", "pid": 4183438, "tid": 31367,
    "ts": 679622573835.743, "dur": 543.914,
    "args": {
      "External id": 537321,"Record function id": 0, "Ev Idx": 2792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622573890.371, "dur": 6.283,
    "args": {
      "External id": 537322,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622573912.201, "dur": 26.585,
    "args": {
      "External id": 537323,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573916.365, "dur": 1.658,
    "args": {
      "External id": 537324,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573919.727, "dur": 0.362,
    "args": {
      "External id": 537325,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573922.018, "dur": 0.435,
    "args": {
      "External id": 537326,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573923.160, "dur": 2.494,
    "args": {
      "External id": 537327,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573926.423, "dur": 0.569,
    "args": {
      "External id": 537328,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573928.422, "dur": 0.567,
    "args": {
      "External id": 537329,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573929.725, "dur": 0.243,
    "args": {
      "External id": 537330,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573931.457, "dur": 1.341,
    "args": {
      "External id": 537331,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622573934.151, "dur": 0.524,
    "args": {
      "External id": 537332,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622573950.843, "dur": 30.654,
    "args": {
      "External id": 537333,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622574016.286, "dur": 122.409,
    "args": {
      "External id": 537334,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 2805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622574025.733, "dur": 4.101,
    "args": {
      "External id": 537335,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622574035.258, "dur": 9.346,
    "args": {
      "External id": 537336,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622574039.213, "dur": 4.998,
    "args": {
      "External id": 537337,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 2808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574042.456, "dur": 0.492,
    "args": {
      "External id": 537338,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622574050.905, "dur": 29.693,
    "args": {
      "External id": 537339,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574052.617, "dur": 2.737,
    "args": {
      "External id": 537340,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574056.825, "dur": 4.223,
    "args": {
      "External id": 537341,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574062.224, "dur": 0.439,
    "args": {
      "External id": 537342,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574063.960, "dur": 0.445,
    "args": {
      "External id": 537343,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574066.658, "dur": 0.385,
    "args": {
      "External id": 537344,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574068.269, "dur": 0.375,
    "args": {
      "External id": 537345,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574070.130, "dur": 0.470,
    "args": {
      "External id": 537346,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574072.773, "dur": 0.478,
    "args": {
      "External id": 537347,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622574074.511, "dur": 2.133,
    "args": {
      "External id": 537348,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622574091.259, "dur": 38.018,
    "args": {
      "External id": 537349,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622574188.507, "dur": 120.396,
    "args": {
      "External id": 537350,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 2821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622574214.008, "dur": 91.255,
    "args": {
      "External id": 537351,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2822, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622574223.820, "dur": 77.426,
    "args": {
      "External id": 537352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 2823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622574323.388, "dur": 1.983,
    "args": {
      "External id": 537353,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2824, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622574410.118, "dur": 2191.754,
    "args": {
      "External id": 537354,"Sequence number": 6319440, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2825
    }
  },
  {
    "ph": "f", "id": 188, "pid": 4183438, "tid": 31367, "ts": 679622574410.118,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622574520.694, "dur": 100.247,
    "args": {
      "External id": 537355,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 2826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622574700.656, "dur": 39.641,
    "args": {
      "External id": 537356,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 2827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622574761.556, "dur": 50.031,
    "args": {
      "External id": 537357,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 2828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622574821.310, "dur": 27.463,
    "args": {
      "External id": 537358,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622574854.872, "dur": 34.952,
    "args": {
      "External id": 537359,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622574896.917, "dur": 21.759,
    "args": {
      "External id": 537360,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622574927.806, "dur": 29.660,
    "args": {
      "External id": 537361,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622574980.944, "dur": 23.264,
    "args": {
      "External id": 537362,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 2833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622575021.527, "dur": 28.999,
    "args": {
      "External id": 537363,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622575068.464, "dur": 18.295,
    "args": {
      "External id": 537364,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622575114.675, "dur": 20.815,
    "args": {
      "External id": 537365,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575147.377, "dur": 45.586,
    "args": {
      "External id": 537366,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575199.352, "dur": 40.050,
    "args": {
      "External id": 537367,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622575287.587, "dur": 211.411,
    "args": {
      "External id": 537368,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622575393.031, "dur": 8.166,
    "args": {
      "External id": 537369,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622575403.936, "dur": 4.227,
    "args": {
      "External id": 537370,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622575534.676, "dur": 26.399,
    "args": {
      "External id": 537371,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622575572.600, "dur": 13.795,
    "args": {
      "External id": 537372,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575595.834, "dur": 94.402,
    "args": {
      "External id": 537373,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575705.846, "dur": 56.975,
    "args": {
      "External id": 537374,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575776.136, "dur": 45.196,
    "args": {
      "External id": 537375,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575836.884, "dur": 55.351,
    "args": {
      "External id": 537376,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575905.193, "dur": 36.639,
    "args": {
      "External id": 537377,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622575955.533, "dur": 221.739,
    "args": {
      "External id": 537378,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622576383.445, "dur": 35.483,
    "args": {
      "External id": 537379,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 2850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622576445.361, "dur": 26.998,
    "args": {
      "External id": 537380,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622576490.989, "dur": 16.849,
    "args": {
      "External id": 537381,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622576524.802, "dur": 14.725,
    "args": {
      "External id": 537382,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622576553.435, "dur": 15.668,
    "args": {
      "External id": 537383,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 2854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576646.536, "dur": 52.195,
    "args": {
      "External id": 537384,"Record function id": 0, "Ev Idx": 2855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576649.675, "dur": 47.248,
    "args": {
      "External id": 537385,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576687.464, "dur": 8.129,
    "args": {
      "External id": 537386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576688.984, "dur": 6.309,
    "args": {
      "External id": 537387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576705.038, "dur": 5.555,
    "args": {
      "External id": 537388,"Record function id": 0, "Ev Idx": 2859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576707.177, "dur": 2.957,
    "args": {
      "External id": 537389,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576708.114, "dur": 1.513,
    "args": {
      "External id": 537390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576708.676, "dur": 0.883,
    "args": {
      "External id": 537391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576714.011, "dur": 4.302,
    "args": {
      "External id": 537392,"Record function id": 0, "Ev Idx": 2863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576715.383, "dur": 2.519,
    "args": {
      "External id": 537393,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576716.015, "dur": 1.455,
    "args": {
      "External id": 537394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576716.518, "dur": 0.841,
    "args": {
      "External id": 537395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 2866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576721.453, "dur": 4.317,
    "args": {
      "External id": 537396,"Record function id": 0, "Ev Idx": 2867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576723.026, "dur": 2.316,
    "args": {
      "External id": 537397,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576723.722, "dur": 1.188,
    "args": {
      "External id": 537398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576724.206, "dur": 0.597,
    "args": {
      "External id": 537399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 2870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576728.943, "dur": 3.673,
    "args": {
      "External id": 537400,"Record function id": 0, "Ev Idx": 2871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576730.150, "dur": 2.055,
    "args": {
      "External id": 537401,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576730.623, "dur": 1.056,
    "args": {
      "External id": 537402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576731.007, "dur": 0.603,
    "args": {
      "External id": 537403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576736.126, "dur": 3.629,
    "args": {
      "External id": 537404,"Record function id": 0, "Ev Idx": 2875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576737.207, "dur": 2.113,
    "args": {
      "External id": 537405,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576737.686, "dur": 1.177,
    "args": {
      "External id": 537406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576738.225, "dur": 0.545,
    "args": {
      "External id": 537407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576743.342, "dur": 3.417,
    "args": {
      "External id": 537408,"Record function id": 0, "Ev Idx": 2879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576744.485, "dur": 1.879,
    "args": {
      "External id": 537409,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576744.987, "dur": 0.947,
    "args": {
      "External id": 537410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576745.352, "dur": 0.487,
    "args": {
      "External id": 537411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576750.104, "dur": 5.819,
    "args": {
      "External id": 537412,"Record function id": 0, "Ev Idx": 2883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576751.249, "dur": 4.236,
    "args": {
      "External id": 537413,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576751.734, "dur": 3.307,
    "args": {
      "External id": 537414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576752.373, "dur": 2.592,
    "args": {
      "External id": 537415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 2886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576759.026, "dur": 3.849,
    "args": {
      "External id": 537416,"Record function id": 0, "Ev Idx": 2887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622576760.310, "dur": 2.190,
    "args": {
      "External id": 537417,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576760.776, "dur": 1.298,
    "args": {
      "External id": 537418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622576761.340, "dur": 0.625,
    "args": {
      "External id": 537419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 2890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622576767.064, "dur": 17875.704,
    "args": {
      "External id": 537420,"Record function id": 0, "Sequence number": 6319439, "Fwd thread id": 1, "Ev Idx": 2891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622576768.251, "dur": 17865.859,
    "args": {
      "External id": 537421,"Sequence number": 6319439, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2892
    }
  },
  {
    "ph": "f", "id": 189, "pid": 4183438, "tid": 31367, "ts": 679622576768.251,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.11)", "pid": 4183438, "tid": 31367,
    "ts": 679622576799.151, "dur": 44.272,
    "args": {
      "External id": 537422,"Record function id": 0, "Ev Idx": 2893
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.11)", "pid": 4183438, "tid": 31367,
    "ts": 679622576851.782, "dur": 67.309,
    "args": {
      "External id": 537423,"Record function id": 0, "Ev Idx": 2894
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.11)", "pid": 4183438, "tid": 31367,
    "ts": 679622576926.322, "dur": 17700.382,
    "args": {
      "External id": 537424,"Record function id": 0, "Ev Idx": 2895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622577017.922, "dur": 8.253,
    "args": {
      "External id": 537425,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622577035.573, "dur": 4.970,
    "args": {
      "External id": 537426,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622577054.158, "dur": 16804.692,
    "args": {
      "External id": 537427,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622577067.724, "dur": 16779.038,
    "args": {
      "External id": 537428,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 2899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622577094.140, "dur": 29.137,
    "args": {
      "External id": 537429,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622577128.136, "dur": 16676.108,
    "args": {
      "External id": 537430,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 2901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622577130.620, "dur": 16672.540,
    "args": {
      "External id": 537431,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 2902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622577134.676, "dur": 6.582,
    "args": {
      "External id": 537432,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622577142.831, "dur": 16654.765,
    "args": {
      "External id": 537433,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 2904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622593965.024, "dur": 12.453,
    "args": {
      "External id": 537434,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 2905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622593968.576, "dur": 8.548,
    "args": {
      "External id": 537435,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622594007.539, "dur": 324.620,
    "args": {
      "External id": 537436,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 2907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622594034.973, "dur": 292.288,
    "args": {
      "External id": 537437,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2908, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622594048.082, "dur": 273.532,
    "args": {
      "External id": 537438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 2909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622594351.225, "dur": 2.391,
    "args": {
      "External id": 537439,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2910, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594414.057, "dur": 6.678,
    "args": {
      "External id": 537440,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594466.700, "dur": 1.605,
    "args": {
      "External id": 537441,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594484.559, "dur": 1.615,
    "args": {
      "External id": 537442,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594497.392, "dur": 1.036,
    "args": {
      "External id": 537443,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594510.323, "dur": 0.721,
    "args": {
      "External id": 537444,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594522.891, "dur": 0.644,
    "args": {
      "External id": 537445,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594534.234, "dur": 0.832,
    "args": {
      "External id": 537446,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594546.706, "dur": 4.799,
    "args": {
      "External id": 537447,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594561.280, "dur": 0.711,
    "args": {
      "External id": 537448,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622594695.674, "dur": 3122.145,
    "args": {
      "External id": 537449,"Record function id": 0, "Ev Idx": 2920
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.10)", "pid": 4183438, "tid": 31367,
    "ts": 679622594717.420, "dur": 1028.781,
    "args": {
      "External id": 537450,"Record function id": 0, "Ev Idx": 2921
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.10)", "pid": 4183438, "tid": 31367,
    "ts": 679622594735.633, "dur": 306.823,
    "args": {
      "External id": 537451,"Record function id": 0, "Ev Idx": 2922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594815.767, "dur": 5.081,
    "args": {
      "External id": 537452,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 2923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594824.324, "dur": 0.840,
    "args": {
      "External id": 537453,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594826.781, "dur": 0.936,
    "args": {
      "External id": 537454,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594829.711, "dur": 0.694,
    "args": {
      "External id": 537455,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594832.521, "dur": 0.789,
    "args": {
      "External id": 537456,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594834.850, "dur": 0.822,
    "args": {
      "External id": 537457,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 2928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594837.386, "dur": 1.544,
    "args": {
      "External id": 537458,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 2929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594840.466, "dur": 3.324,
    "args": {
      "External id": 537459,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594845.475, "dur": 0.622,
    "args": {
      "External id": 537460,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622594847.914, "dur": 0.637,
    "args": {
      "External id": 537461,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 2932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622594866.775, "dur": 140.965,
    "args": {
      "External id": 537462,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622594882.899, "dur": 120.218,
    "args": {
      "External id": 537463,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 2934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622594896.088, "dur": 12.635,
    "args": {
      "External id": 537464,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622594911.981, "dur": 64.474,
    "args": {
      "External id": 537465,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 2936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622594914.558, "dur": 61.507,
    "args": {
      "External id": 537466,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 2937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622594918.658, "dur": 6.534,
    "args": {
      "External id": 537467,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622594926.905, "dur": 48.328,
    "args": {
      "External id": 537468,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 2939
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.9", "pid": 4183438, "tid": 31367,
    "ts": 679622595144.755, "dur": 594.267,
    "args": {
      "External id": 537469,"Record function id": 0, "Ev Idx": 2940
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.9)", "pid": 4183438, "tid": 31367,
    "ts": 679622595166.151, "dur": 559.465,
    "args": {
      "External id": 537470,"Record function id": 0, "Ev Idx": 2941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622595226.911, "dur": 6.011,
    "args": {
      "External id": 537471,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622595247.725, "dur": 33.936,
    "args": {
      "External id": 537472,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595253.038, "dur": 2.535,
    "args": {
      "External id": 537473,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595257.592, "dur": 0.368,
    "args": {
      "External id": 537474,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595259.246, "dur": 2.245,
    "args": {
      "External id": 537475,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595263.826, "dur": 0.316,
    "args": {
      "External id": 537476,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595265.740, "dur": 0.726,
    "args": {
      "External id": 537477,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595267.431, "dur": 0.247,
    "args": {
      "External id": 537478,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595274.086, "dur": 0.429,
    "args": {
      "External id": 537479,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595275.573, "dur": 0.277,
    "args": {
      "External id": 537480,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595276.600, "dur": 1.000,
    "args": {
      "External id": 537481,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622595291.350, "dur": 32.208,
    "args": {
      "External id": 537482,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622595358.252, "dur": 99.581,
    "args": {
      "External id": 537483,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 2954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622595368.085, "dur": 4.239,
    "args": {
      "External id": 537484,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622595378.036, "dur": 11.327,
    "args": {
      "External id": 537485,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622595382.320, "dur": 6.647,
    "args": {
      "External id": 537486,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 2957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595385.617, "dur": 2.180,
    "args": {
      "External id": 537487,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 2958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622595395.848, "dur": 24.325,
    "args": {
      "External id": 537488,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 2959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595397.657, "dur": 0.469,
    "args": {
      "External id": 537489,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595399.512, "dur": 1.468,
    "args": {
      "External id": 537490,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595403.015, "dur": 0.433,
    "args": {
      "External id": 537491,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595404.650, "dur": 0.776,
    "args": {
      "External id": 537492,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595408.092, "dur": 0.281,
    "args": {
      "External id": 537493,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595409.363, "dur": 0.445,
    "args": {
      "External id": 537494,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595411.384, "dur": 0.420,
    "args": {
      "External id": 537495,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595413.561, "dur": 1.866,
    "args": {
      "External id": 537496,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622595416.857, "dur": 0.383,
    "args": {
      "External id": 537497,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 2968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622595432.739, "dur": 17.991,
    "args": {
      "External id": 537498,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 2969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622595502.007, "dur": 114.133,
    "args": {
      "External id": 537499,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 2970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622595526.431, "dur": 86.136,
    "args": {
      "External id": 537500,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2971, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622595536.579, "dur": 71.746,
    "args": {
      "External id": 537501,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 2972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622595630.714, "dur": 1.900,
    "args": {
      "External id": 537502,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2973, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622595754.305, "dur": 2039.951,
    "args": {
      "External id": 537503,"Sequence number": 6319438, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 2974
    }
  },
  {
    "ph": "f", "id": 190, "pid": 4183438, "tid": 31367, "ts": 679622595754.305,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622595865.448, "dur": 100.820,
    "args": {
      "External id": 537504,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 2975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622596004.290, "dur": 36.505,
    "args": {
      "External id": 537505,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 2976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622596058.422, "dur": 56.397,
    "args": {
      "External id": 537506,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 2977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596127.979, "dur": 30.241,
    "args": {
      "External id": 537507,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596163.917, "dur": 34.612,
    "args": {
      "External id": 537508,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596205.437, "dur": 23.016,
    "args": {
      "External id": 537509,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 2980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596236.985, "dur": 30.197,
    "args": {
      "External id": 537510,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 2981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622596289.955, "dur": 23.788,
    "args": {
      "External id": 537511,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 2982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622596333.460, "dur": 27.186,
    "args": {
      "External id": 537512,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622596380.385, "dur": 18.516,
    "args": {
      "External id": 537513,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 2984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622596411.351, "dur": 13.943,
    "args": {
      "External id": 537514,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 2985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596435.333, "dur": 36.575,
    "args": {
      "External id": 537515,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596475.252, "dur": 29.855,
    "args": {
      "External id": 537516,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622596531.929, "dur": 205.616,
    "args": {
      "External id": 537517,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622596604.453, "dur": 5.823,
    "args": {
      "External id": 537518,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622596611.789, "dur": 5.244,
    "args": {
      "External id": 537519,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622596769.852, "dur": 24.272,
    "args": {
      "External id": 537520,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622596805.120, "dur": 12.832,
    "args": {
      "External id": 537521,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596828.166, "dur": 51.280,
    "args": {
      "External id": 537522,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596884.743, "dur": 33.544,
    "args": {
      "External id": 537523,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596925.188, "dur": 28.050,
    "args": {
      "External id": 537524,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596957.870, "dur": 28.577,
    "args": {
      "External id": 537525,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622596996.309, "dur": 34.293,
    "args": {
      "External id": 537526,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 2997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622597040.017, "dur": 41.492,
    "args": {
      "External id": 537527,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 2998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622597121.963, "dur": 33.897,
    "args": {
      "External id": 537528,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 2999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622597198.289, "dur": 43.127,
    "args": {
      "External id": 537529,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622597267.091, "dur": 25.789,
    "args": {
      "External id": 537530,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622597319.696, "dur": 188.431,
    "args": {
      "External id": 537531,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622597525.308, "dur": 222.912,
    "args": {
      "External id": 537532,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 3003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597842.955, "dur": 15.972,
    "args": {
      "External id": 537533,"Record function id": 0, "Ev Idx": 3004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597846.290, "dur": 11.713,
    "args": {
      "External id": 537534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597851.116, "dur": 5.937,
    "args": {
      "External id": 537535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597852.550, "dur": 4.426,
    "args": {
      "External id": 537536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597862.796, "dur": 5.054,
    "args": {
      "External id": 537537,"Record function id": 0, "Ev Idx": 3008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597864.323, "dur": 3.117,
    "args": {
      "External id": 537538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597865.250, "dur": 1.773,
    "args": {
      "External id": 537539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597865.879, "dur": 1.046,
    "args": {
      "External id": 537540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597871.030, "dur": 3.557,
    "args": {
      "External id": 537541,"Record function id": 0, "Ev Idx": 3012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597872.073, "dur": 2.151,
    "args": {
      "External id": 537542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597872.714, "dur": 1.113,
    "args": {
      "External id": 537543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597873.139, "dur": 0.605,
    "args": {
      "External id": 537544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597877.643, "dur": 4.284,
    "args": {
      "External id": 537545,"Record function id": 0, "Ev Idx": 3016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597879.129, "dur": 2.434,
    "args": {
      "External id": 537546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597879.823, "dur": 1.337,
    "args": {
      "External id": 537547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597880.574, "dur": 0.522,
    "args": {
      "External id": 537548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597884.968, "dur": 4.240,
    "args": {
      "External id": 537549,"Record function id": 0, "Ev Idx": 3020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597886.188, "dur": 2.655,
    "args": {
      "External id": 537550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597886.766, "dur": 1.659,
    "args": {
      "External id": 537551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597887.471, "dur": 0.889,
    "args": {
      "External id": 537552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597892.331, "dur": 3.702,
    "args": {
      "External id": 537553,"Record function id": 0, "Ev Idx": 3024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597893.438, "dur": 2.224,
    "args": {
      "External id": 537554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597893.915, "dur": 1.352,
    "args": {
      "External id": 537555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597894.593, "dur": 0.609,
    "args": {
      "External id": 537556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597899.049, "dur": 5.696,
    "args": {
      "External id": 537557,"Record function id": 0, "Ev Idx": 3028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597900.044, "dur": 4.296,
    "args": {
      "External id": 537558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597900.518, "dur": 3.438,
    "args": {
      "External id": 537559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597901.080, "dur": 2.814,
    "args": {
      "External id": 537560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597907.987, "dur": 3.798,
    "args": {
      "External id": 537561,"Record function id": 0, "Ev Idx": 3032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597908.935, "dur": 2.445,
    "args": {
      "External id": 537562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597909.423, "dur": 1.578,
    "args": {
      "External id": 537563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597910.077, "dur": 0.860,
    "args": {
      "External id": 537564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597915.434, "dur": 3.356,
    "args": {
      "External id": 537565,"Record function id": 0, "Ev Idx": 3036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622597916.302, "dur": 2.113,
    "args": {
      "External id": 537566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597916.764, "dur": 1.265,
    "args": {
      "External id": 537567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622597917.418, "dur": 0.544,
    "args": {
      "External id": 537568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622597922.866, "dur": 15354.712,
    "args": {
      "External id": 537569,"Record function id": 0, "Sequence number": 6319437, "Fwd thread id": 1, "Ev Idx": 3040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622597924.136, "dur": 15344.844,
    "args": {
      "External id": 537570,"Sequence number": 6319437, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3041
    }
  },
  {
    "ph": "f", "id": 191, "pid": 4183438, "tid": 31367, "ts": 679622597924.136,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.10)", "pid": 4183438, "tid": 31367,
    "ts": 679622597957.042, "dur": 41.600,
    "args": {
      "External id": 537571,"Record function id": 0, "Ev Idx": 3042
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.10)", "pid": 4183438, "tid": 31367,
    "ts": 679622598006.827, "dur": 66.457,
    "args": {
      "External id": 537572,"Record function id": 0, "Ev Idx": 3043
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.10)", "pid": 4183438, "tid": 31367,
    "ts": 679622598080.111, "dur": 15180.904,
    "args": {
      "External id": 537573,"Record function id": 0, "Ev Idx": 3044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622598192.056, "dur": 8.565,
    "args": {
      "External id": 537574,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622598211.366, "dur": 5.362,
    "args": {
      "External id": 537575,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622598230.823, "dur": 14286.018,
    "args": {
      "External id": 537576,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622598251.082, "dur": 14254.889,
    "args": {
      "External id": 537577,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622598278.414, "dur": 14.332,
    "args": {
      "External id": 537578,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622598296.883, "dur": 14167.558,
    "args": {
      "External id": 537579,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 3050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622598299.936, "dur": 14163.470,
    "args": {
      "External id": 537580,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 3051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622598303.919, "dur": 6.280,
    "args": {
      "External id": 537581,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622598311.881, "dur": 14146.863,
    "args": {
      "External id": 537582,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 3053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622612624.562, "dur": 11.887,
    "args": {
      "External id": 537583,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 3054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622612628.273, "dur": 7.801,
    "args": {
      "External id": 537584,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622612696.435, "dur": 260.112,
    "args": {
      "External id": 537585,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 3056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622612728.042, "dur": 223.962,
    "args": {
      "External id": 537586,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3057, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622612740.774, "dur": 205.948,
    "args": {
      "External id": 537587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 3058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622612973.220, "dur": 2.099,
    "args": {
      "External id": 537588,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3059, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613032.261, "dur": 8.531,
    "args": {
      "External id": 537589,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613084.475, "dur": 1.476,
    "args": {
      "External id": 537590,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613118.070, "dur": 1.656,
    "args": {
      "External id": 537591,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613134.651, "dur": 0.759,
    "args": {
      "External id": 537592,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613146.663, "dur": 2.545,
    "args": {
      "External id": 537593,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613158.296, "dur": 1.060,
    "args": {
      "External id": 537594,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613170.174, "dur": 0.828,
    "args": {
      "External id": 537595,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613181.309, "dur": 1.565,
    "args": {
      "External id": 537596,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613192.646, "dur": 2.580,
    "args": {
      "External id": 537597,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622613292.874, "dur": 3130.322,
    "args": {
      "External id": 537598,"Record function id": 0, "Ev Idx": 3069
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.9)", "pid": 4183438, "tid": 31367,
    "ts": 679622613315.993, "dur": 1032.636,
    "args": {
      "External id": 537599,"Record function id": 0, "Ev Idx": 3070
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.9)", "pid": 4183438, "tid": 31367,
    "ts": 679622613330.995, "dur": 312.443,
    "args": {
      "External id": 537600,"Record function id": 0, "Ev Idx": 3071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613424.270, "dur": 4.292,
    "args": {
      "External id": 537601,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613431.964, "dur": 0.937,
    "args": {
      "External id": 537602,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613434.658, "dur": 0.916,
    "args": {
      "External id": 537603,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613437.107, "dur": 0.830,
    "args": {
      "External id": 537604,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613439.359, "dur": 0.678,
    "args": {
      "External id": 537605,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613441.562, "dur": 0.831,
    "args": {
      "External id": 537606,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613444.088, "dur": 6.267,
    "args": {
      "External id": 537607,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613451.982, "dur": 0.695,
    "args": {
      "External id": 537608,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613453.957, "dur": 0.528,
    "args": {
      "External id": 537609,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622613456.015, "dur": 0.725,
    "args": {
      "External id": 537610,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622613474.142, "dur": 141.270,
    "args": {
      "External id": 537611,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622613489.424, "dur": 121.893,
    "args": {
      "External id": 537612,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622613506.306, "dur": 13.666,
    "args": {
      "External id": 537613,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622613522.638, "dur": 62.956,
    "args": {
      "External id": 537614,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 3085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622613525.275, "dur": 59.964,
    "args": {
      "External id": 537615,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 3086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613528.913, "dur": 6.101,
    "args": {
      "External id": 537616,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622613536.646, "dur": 47.734,
    "args": {
      "External id": 537617,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 3088
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.8", "pid": 4183438, "tid": 31367,
    "ts": 679622613768.297, "dur": 573.050,
    "args": {
      "External id": 537618,"Record function id": 0, "Ev Idx": 3089
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.8)", "pid": 4183438, "tid": 31367,
    "ts": 679622613785.421, "dur": 543.856,
    "args": {
      "External id": 537619,"Record function id": 0, "Ev Idx": 3090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622613842.012, "dur": 5.586,
    "args": {
      "External id": 537620,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622613863.070, "dur": 28.779,
    "args": {
      "External id": 537621,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613867.916, "dur": 1.496,
    "args": {
      "External id": 537622,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613871.803, "dur": 2.338,
    "args": {
      "External id": 537623,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613875.333, "dur": 0.277,
    "args": {
      "External id": 537624,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613877.515, "dur": 0.362,
    "args": {
      "External id": 537625,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613879.065, "dur": 0.275,
    "args": {
      "External id": 537626,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613880.381, "dur": 1.252,
    "args": {
      "External id": 537627,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613882.976, "dur": 0.340,
    "args": {
      "External id": 537628,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613884.203, "dur": 0.315,
    "args": {
      "External id": 537629,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613887.448, "dur": 0.323,
    "args": {
      "External id": 537630,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622613901.200, "dur": 31.436,
    "args": {
      "External id": 537631,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622613964.044, "dur": 99.548,
    "args": {
      "External id": 537632,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 3103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622613973.681, "dur": 5.421,
    "args": {
      "External id": 537633,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622613984.440, "dur": 9.310,
    "args": {
      "External id": 537634,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622613988.419, "dur": 4.932,
    "args": {
      "External id": 537635,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 3106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622613991.727, "dur": 0.541,
    "args": {
      "External id": 537636,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622614000.022, "dur": 24.418,
    "args": {
      "External id": 537637,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614001.669, "dur": 0.414,
    "args": {
      "External id": 537638,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614004.137, "dur": 0.522,
    "args": {
      "External id": 537639,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614006.672, "dur": 0.429,
    "args": {
      "External id": 537640,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614008.410, "dur": 0.418,
    "args": {
      "External id": 537641,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614009.945, "dur": 0.296,
    "args": {
      "External id": 537642,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614012.629, "dur": 0.401,
    "args": {
      "External id": 537643,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614014.160, "dur": 2.395,
    "args": {
      "External id": 537644,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614017.787, "dur": 1.735,
    "args": {
      "External id": 537645,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622614020.801, "dur": 0.374,
    "args": {
      "External id": 537646,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622614035.515, "dur": 20.784,
    "args": {
      "External id": 537647,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622614134.019, "dur": 117.247,
    "args": {
      "External id": 537648,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 3119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622614159.384, "dur": 88.262,
    "args": {
      "External id": 537649,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3120, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622614170.438, "dur": 73.081,
    "args": {
      "External id": 537650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 3121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622614269.622, "dur": 1.769,
    "args": {
      "External id": 537651,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3122, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622614355.452, "dur": 2044.036,
    "args": {
      "External id": 537652,"Sequence number": 6319436, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3123
    }
  },
  {
    "ph": "f", "id": 192, "pid": 4183438, "tid": 31367, "ts": 679622614355.452,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622614472.347, "dur": 98.803,
    "args": {
      "External id": 537653,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 3124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622614609.373, "dur": 35.176,
    "args": {
      "External id": 537654,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 3125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622614706.725, "dur": 50.481,
    "args": {
      "External id": 537655,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 3126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622614766.783, "dur": 27.720,
    "args": {
      "External id": 537656,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622614800.600, "dur": 34.273,
    "args": {
      "External id": 537657,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622614843.297, "dur": 21.514,
    "args": {
      "External id": 537658,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622614872.286, "dur": 28.631,
    "args": {
      "External id": 537659,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622614924.637, "dur": 24.622,
    "args": {
      "External id": 537660,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 3131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622614966.768, "dur": 29.532,
    "args": {
      "External id": 537661,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622615014.691, "dur": 20.200,
    "args": {
      "External id": 537662,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622615047.190, "dur": 14.728,
    "args": {
      "External id": 537663,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615071.959, "dur": 51.843,
    "args": {
      "External id": 537664,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615128.866, "dur": 34.106,
    "args": {
      "External id": 537665,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622615191.425, "dur": 169.126,
    "args": {
      "External id": 537666,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622615267.972, "dur": 6.267,
    "args": {
      "External id": 537667,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622615276.083, "dur": 5.255,
    "args": {
      "External id": 537668,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622615390.115, "dur": 22.392,
    "args": {
      "External id": 537669,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622615426.418, "dur": 13.843,
    "args": {
      "External id": 537670,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615447.944, "dur": 42.335,
    "args": {
      "External id": 537671,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615495.933, "dur": 30.549,
    "args": {
      "External id": 537672,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615533.765, "dur": 28.161,
    "args": {
      "External id": 537673,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615568.314, "dur": 31.341,
    "args": {
      "External id": 537674,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615608.759, "dur": 34.154,
    "args": {
      "External id": 537675,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622615653.342, "dur": 77.715,
    "args": {
      "External id": 537676,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622615764.683, "dur": 41.656,
    "args": {
      "External id": 537677,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 3148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622615842.718, "dur": 36.642,
    "args": {
      "External id": 537678,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622615899.854, "dur": 23.351,
    "args": {
      "External id": 537679,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622616126.779, "dur": 22.577,
    "args": {
      "External id": 537680,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622616338.461, "dur": 21.932,
    "args": {
      "External id": 537681,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 3152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616447.036, "dur": 16.827,
    "args": {
      "External id": 537682,"Record function id": 0, "Ev Idx": 3153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616450.281, "dur": 12.617,
    "args": {
      "External id": 537683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616455.396, "dur": 6.526,
    "args": {
      "External id": 537684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616457.238, "dur": 4.590,
    "args": {
      "External id": 537685,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616467.644, "dur": 4.778,
    "args": {
      "External id": 537686,"Record function id": 0, "Ev Idx": 3157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616469.254, "dur": 2.771,
    "args": {
      "External id": 537687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616470.145, "dur": 1.442,
    "args": {
      "External id": 537688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616470.615, "dur": 0.888,
    "args": {
      "External id": 537689,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616475.428, "dur": 4.193,
    "args": {
      "External id": 537690,"Record function id": 0, "Ev Idx": 3161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616476.868, "dur": 2.400,
    "args": {
      "External id": 537691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616477.352, "dur": 1.500,
    "args": {
      "External id": 537692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616477.873, "dur": 0.897,
    "args": {
      "External id": 537693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616482.601, "dur": 6.054,
    "args": {
      "External id": 537694,"Record function id": 0, "Ev Idx": 3165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616483.729, "dur": 4.539,
    "args": {
      "External id": 537695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616484.367, "dur": 3.509,
    "args": {
      "External id": 537696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616485.093, "dur": 2.720,
    "args": {
      "External id": 537697,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616491.528, "dur": 4.209,
    "args": {
      "External id": 537698,"Record function id": 0, "Ev Idx": 3169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616492.682, "dur": 2.677,
    "args": {
      "External id": 537699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616493.170, "dur": 1.778,
    "args": {
      "External id": 537700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616493.939, "dur": 0.939,
    "args": {
      "External id": 537701,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616498.648, "dur": 3.582,
    "args": {
      "External id": 537702,"Record function id": 0, "Ev Idx": 3173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616499.746, "dur": 2.095,
    "args": {
      "External id": 537703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616500.233, "dur": 1.208,
    "args": {
      "External id": 537704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616500.711, "dur": 0.664,
    "args": {
      "External id": 537705,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616505.391, "dur": 3.934,
    "args": {
      "External id": 537706,"Record function id": 0, "Ev Idx": 3177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616506.442, "dur": 2.514,
    "args": {
      "External id": 537707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616507.120, "dur": 1.434,
    "args": {
      "External id": 537708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616507.724, "dur": 0.756,
    "args": {
      "External id": 537709,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616512.539, "dur": 3.849,
    "args": {
      "External id": 537710,"Record function id": 0, "Ev Idx": 3181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616513.685, "dur": 2.338,
    "args": {
      "External id": 537711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616514.294, "dur": 1.309,
    "args": {
      "External id": 537712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616514.974, "dur": 0.556,
    "args": {
      "External id": 537713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616519.634, "dur": 3.572,
    "args": {
      "External id": 537714,"Record function id": 0, "Ev Idx": 3185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622616520.769, "dur": 2.031,
    "args": {
      "External id": 537715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616521.236, "dur": 1.171,
    "args": {
      "External id": 537716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622616521.563, "dur": 0.780,
    "args": {
      "External id": 537717,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622616526.954, "dur": 15222.842,
    "args": {
      "External id": 537718,"Record function id": 0, "Sequence number": 6319435, "Fwd thread id": 1, "Ev Idx": 3189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622616528.298, "dur": 15212.472,
    "args": {
      "External id": 537719,"Sequence number": 6319435, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3190
    }
  },
  {
    "ph": "f", "id": 193, "pid": 4183438, "tid": 31367, "ts": 679622616528.298,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.9)", "pid": 4183438, "tid": 31367,
    "ts": 679622616560.664, "dur": 38.746,
    "args": {
      "External id": 537720,"Record function id": 0, "Ev Idx": 3191
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.9)", "pid": 4183438, "tid": 31367,
    "ts": 679622616607.314, "dur": 101.394,
    "args": {
      "External id": 537721,"Record function id": 0, "Ev Idx": 3192
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.9)", "pid": 4183438, "tid": 31367,
    "ts": 679622616717.691, "dur": 15014.879,
    "args": {
      "External id": 537722,"Record function id": 0, "Ev Idx": 3193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622616810.265, "dur": 8.606,
    "args": {
      "External id": 537723,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622616829.908, "dur": 5.453,
    "args": {
      "External id": 537724,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622616850.895, "dur": 14133.156,
    "args": {
      "External id": 537725,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622616864.250, "dur": 14109.339,
    "args": {
      "External id": 537726,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622616890.906, "dur": 16.313,
    "args": {
      "External id": 537727,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622616911.213, "dur": 14017.232,
    "args": {
      "External id": 537728,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 3199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622616913.601, "dur": 14013.894,
    "args": {
      "External id": 537729,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 3200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622616917.825, "dur": 4.988,
    "args": {
      "External id": 537730,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622616924.347, "dur": 13998.953,
    "args": {
      "External id": 537731,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 3202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622631089.634, "dur": 24.645,
    "args": {
      "External id": 537732,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 3203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622631093.148, "dur": 20.387,
    "args": {
      "External id": 537733,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622631148.764, "dur": 274.090,
    "args": {
      "External id": 537734,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 3205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622631181.494, "dur": 237.127,
    "args": {
      "External id": 537735,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3206, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622631198.105, "dur": 214.874,
    "args": {
      "External id": 537736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 3207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622631439.297, "dur": 1.964,
    "args": {
      "External id": 537737,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3208, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631497.649, "dur": 6.305,
    "args": {
      "External id": 537738,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631547.275, "dur": 1.482,
    "args": {
      "External id": 537739,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631563.488, "dur": 1.336,
    "args": {
      "External id": 537740,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631578.599, "dur": 0.996,
    "args": {
      "External id": 537741,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631591.164, "dur": 0.810,
    "args": {
      "External id": 537742,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631601.088, "dur": 0.890,
    "args": {
      "External id": 537743,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631612.124, "dur": 0.883,
    "args": {
      "External id": 537744,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631624.651, "dur": 1.865,
    "args": {
      "External id": 537745,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631636.066, "dur": 0.856,
    "args": {
      "External id": 537746,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622631766.371, "dur": 3136.543,
    "args": {
      "External id": 537747,"Record function id": 0, "Ev Idx": 3218
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.8)", "pid": 4183438, "tid": 31367,
    "ts": 679622631785.587, "dur": 1005.037,
    "args": {
      "External id": 537748,"Record function id": 0, "Ev Idx": 3219
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.8)", "pid": 4183438, "tid": 31367,
    "ts": 679622631801.467, "dur": 297.145,
    "args": {
      "External id": 537749,"Record function id": 0, "Ev Idx": 3220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631880.788, "dur": 4.977,
    "args": {
      "External id": 537750,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631889.157, "dur": 0.833,
    "args": {
      "External id": 537751,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631891.706, "dur": 0.850,
    "args": {
      "External id": 537752,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631894.368, "dur": 2.648,
    "args": {
      "External id": 537753,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631898.826, "dur": 0.671,
    "args": {
      "External id": 537754,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631901.031, "dur": 0.793,
    "args": {
      "External id": 537755,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631903.814, "dur": 1.358,
    "args": {
      "External id": 537756,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631907.011, "dur": 0.943,
    "args": {
      "External id": 537757,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631909.584, "dur": 0.615,
    "args": {
      "External id": 537758,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622631911.684, "dur": 0.581,
    "args": {
      "External id": 537759,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622631930.217, "dur": 141.844,
    "args": {
      "External id": 537760,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622631946.697, "dur": 121.385,
    "args": {
      "External id": 537761,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622631959.688, "dur": 14.786,
    "args": {
      "External id": 537762,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622631977.347, "dur": 64.274,
    "args": {
      "External id": 537763,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 3234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622631979.986, "dur": 61.321,
    "args": {
      "External id": 537764,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 3235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622631983.678, "dur": 7.854,
    "args": {
      "External id": 537765,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622631993.273, "dur": 47.272,
    "args": {
      "External id": 537766,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 3237
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.7", "pid": 4183438, "tid": 31367,
    "ts": 679622632199.334, "dur": 583.875,
    "args": {
      "External id": 537767,"Record function id": 0, "Ev Idx": 3238
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.7)", "pid": 4183438, "tid": 31367,
    "ts": 679622632214.804, "dur": 555.545,
    "args": {
      "External id": 537768,"Record function id": 0, "Ev Idx": 3239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622632271.867, "dur": 5.481,
    "args": {
      "External id": 537769,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622632292.442, "dur": 30.788,
    "args": {
      "External id": 537770,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632297.861, "dur": 1.562,
    "args": {
      "External id": 537771,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632302.866, "dur": 0.574,
    "args": {
      "External id": 537772,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632305.075, "dur": 0.710,
    "args": {
      "External id": 537773,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632306.775, "dur": 1.918,
    "args": {
      "External id": 537774,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632310.442, "dur": 0.478,
    "args": {
      "External id": 537775,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632311.755, "dur": 0.480,
    "args": {
      "External id": 537776,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632314.700, "dur": 1.892,
    "args": {
      "External id": 537777,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632317.308, "dur": 0.626,
    "args": {
      "External id": 537778,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632318.993, "dur": 0.325,
    "args": {
      "External id": 537779,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622632332.906, "dur": 35.473,
    "args": {
      "External id": 537780,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622632401.424, "dur": 97.695,
    "args": {
      "External id": 537781,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 3252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622632411.147, "dur": 3.300,
    "args": {
      "External id": 537782,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622632419.542, "dur": 10.479,
    "args": {
      "External id": 537783,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622632423.779, "dur": 5.832,
    "args": {
      "External id": 537784,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 3255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632427.037, "dur": 1.468,
    "args": {
      "External id": 537785,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622632436.371, "dur": 26.256,
    "args": {
      "External id": 537786,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632438.415, "dur": 0.508,
    "args": {
      "External id": 537787,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632443.527, "dur": 0.783,
    "args": {
      "External id": 537788,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632445.325, "dur": 0.292,
    "args": {
      "External id": 537789,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632447.282, "dur": 2.336,
    "args": {
      "External id": 537790,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632450.765, "dur": 0.622,
    "args": {
      "External id": 537791,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632452.464, "dur": 0.875,
    "args": {
      "External id": 537792,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632454.601, "dur": 0.371,
    "args": {
      "External id": 537793,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632455.829, "dur": 0.442,
    "args": {
      "External id": 537794,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622632458.809, "dur": 0.405,
    "args": {
      "External id": 537795,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622632472.364, "dur": 19.491,
    "args": {
      "External id": 537796,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622632541.170, "dur": 155.822,
    "args": {
      "External id": 537797,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 3268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622632561.897, "dur": 130.823,
    "args": {
      "External id": 537798,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3269, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622632571.444, "dur": 78.803,
    "args": {
      "External id": 537799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 3270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622632714.098, "dur": 2.184,
    "args": {
      "External id": 537800,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3271, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622632798.298, "dur": 1846.633,
    "args": {
      "External id": 537801,"Sequence number": 6319434, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3272
    }
  },
  {
    "ph": "f", "id": 194, "pid": 4183438, "tid": 31367, "ts": 679622632798.298,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622632912.210, "dur": 109.629,
    "args": {
      "External id": 537802,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 3273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622633064.792, "dur": 54.313,
    "args": {
      "External id": 537803,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 3274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622633138.732, "dur": 46.672,
    "args": {
      "External id": 537804,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 3275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633195.156, "dur": 26.594,
    "args": {
      "External id": 537805,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633229.732, "dur": 34.740,
    "args": {
      "External id": 537806,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633275.363, "dur": 22.777,
    "args": {
      "External id": 537807,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633305.519, "dur": 30.231,
    "args": {
      "External id": 537808,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622633359.559, "dur": 23.628,
    "args": {
      "External id": 537809,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 3280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622633402.967, "dur": 29.612,
    "args": {
      "External id": 537810,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622633449.186, "dur": 20.422,
    "args": {
      "External id": 537811,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622633481.205, "dur": 15.405,
    "args": {
      "External id": 537812,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633505.000, "dur": 35.871,
    "args": {
      "External id": 537813,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633544.393, "dur": 30.424,
    "args": {
      "External id": 537814,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622633602.412, "dur": 207.268,
    "args": {
      "External id": 537815,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622633716.717, "dur": 7.478,
    "args": {
      "External id": 537816,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622633726.207, "dur": 2.619,
    "args": {
      "External id": 537817,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622633840.212, "dur": 22.426,
    "args": {
      "External id": 537818,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622633873.212, "dur": 14.449,
    "args": {
      "External id": 537819,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633895.586, "dur": 49.393,
    "args": {
      "External id": 537820,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633950.469, "dur": 31.132,
    "args": {
      "External id": 537821,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622633990.716, "dur": 27.969,
    "args": {
      "External id": 537822,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622634023.711, "dur": 27.328,
    "args": {
      "External id": 537823,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622634058.962, "dur": 33.770,
    "args": {
      "External id": 537824,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622634119.283, "dur": 46.584,
    "args": {
      "External id": 537825,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622634190.601, "dur": 30.755,
    "args": {
      "External id": 537826,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 3297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622634246.714, "dur": 51.633,
    "args": {
      "External id": 537827,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622634325.198, "dur": 25.175,
    "args": {
      "External id": 537828,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622634375.691, "dur": 22.491,
    "args": {
      "External id": 537829,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622634415.026, "dur": 24.288,
    "args": {
      "External id": 537830,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 3301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634931.851, "dur": 19.174,
    "args": {
      "External id": 537831,"Record function id": 0, "Ev Idx": 3302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634935.840, "dur": 13.848,
    "args": {
      "External id": 537832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634941.016, "dur": 7.547,
    "args": {
      "External id": 537833,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634942.951, "dur": 5.389,
    "args": {
      "External id": 537834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634956.551, "dur": 7.736,
    "args": {
      "External id": 537835,"Record function id": 0, "Ev Idx": 3306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634959.017, "dur": 4.535,
    "args": {
      "External id": 537836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634960.404, "dur": 2.542,
    "args": {
      "External id": 537837,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634961.347, "dur": 1.408,
    "args": {
      "External id": 537838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634969.230, "dur": 6.567,
    "args": {
      "External id": 537839,"Record function id": 0, "Ev Idx": 3310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634970.642, "dur": 4.719,
    "args": {
      "External id": 537840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634971.468, "dur": 3.449,
    "args": {
      "External id": 537841,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634972.091, "dur": 2.734,
    "args": {
      "External id": 537842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634979.025, "dur": 4.688,
    "args": {
      "External id": 537843,"Record function id": 0, "Ev Idx": 3314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634980.345, "dur": 2.991,
    "args": {
      "External id": 537844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634981.128, "dur": 1.824,
    "args": {
      "External id": 537845,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634982.046, "dur": 0.839,
    "args": {
      "External id": 537846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634986.724, "dur": 3.858,
    "args": {
      "External id": 537847,"Record function id": 0, "Ev Idx": 3318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634987.917, "dur": 2.287,
    "args": {
      "External id": 537848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634988.566, "dur": 1.249,
    "args": {
      "External id": 537849,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634988.988, "dur": 0.761,
    "args": {
      "External id": 537850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634993.686, "dur": 3.913,
    "args": {
      "External id": 537851,"Record function id": 0, "Ev Idx": 3322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622634994.883, "dur": 2.323,
    "args": {
      "External id": 537852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634995.500, "dur": 1.338,
    "args": {
      "External id": 537853,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622634996.096, "dur": 0.677,
    "args": {
      "External id": 537854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622635000.800, "dur": 3.806,
    "args": {
      "External id": 537855,"Record function id": 0, "Ev Idx": 3326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622635002.065, "dur": 2.131,
    "args": {
      "External id": 537856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622635002.496, "dur": 1.292,
    "args": {
      "External id": 537857,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622635003.026, "dur": 0.685,
    "args": {
      "External id": 537858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622635007.518, "dur": 3.911,
    "args": {
      "External id": 537859,"Record function id": 0, "Ev Idx": 3330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622635008.486, "dur": 2.563,
    "args": {
      "External id": 537860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622635008.947, "dur": 1.710,
    "args": {
      "External id": 537861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622635009.774, "dur": 0.808,
    "args": {
      "External id": 537862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622635014.385, "dur": 3.750,
    "args": {
      "External id": 537863,"Record function id": 0, "Ev Idx": 3334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622635015.507, "dur": 2.241,
    "args": {
      "External id": 537864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622635016.089, "dur": 1.272,
    "args": {
      "External id": 537865,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622635016.620, "dur": 0.674,
    "args": {
      "External id": 537866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622635021.835, "dur": 15093.307,
    "args": {
      "External id": 537867,"Record function id": 0, "Sequence number": 6319433, "Fwd thread id": 1, "Ev Idx": 3338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622635023.126, "dur": 15066.992,
    "args": {
      "External id": 537868,"Sequence number": 6319433, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3339
    }
  },
  {
    "ph": "f", "id": 195, "pid": 4183438, "tid": 31367, "ts": 679622635023.126,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.8)", "pid": 4183438, "tid": 31367,
    "ts": 679622635055.674, "dur": 42.820,
    "args": {
      "External id": 537869,"Record function id": 0, "Ev Idx": 3340
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.8)", "pid": 4183438, "tid": 31367,
    "ts": 679622635122.864, "dur": 68.787,
    "args": {
      "External id": 537870,"Record function id": 0, "Ev Idx": 3341
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.8)", "pid": 4183438, "tid": 31367,
    "ts": 679622635198.566, "dur": 14883.423,
    "args": {
      "External id": 537871,"Record function id": 0, "Ev Idx": 3342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622635294.532, "dur": 8.361,
    "args": {
      "External id": 537872,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622635313.325, "dur": 6.875,
    "args": {
      "External id": 537873,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622635334.102, "dur": 13990.852,
    "args": {
      "External id": 537874,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622635347.243, "dur": 13966.552,
    "args": {
      "External id": 537875,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622635370.267, "dur": 13.785,
    "args": {
      "External id": 537876,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622635387.966, "dur": 13882.853,
    "args": {
      "External id": 537877,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 3348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622635391.517, "dur": 13878.375,
    "args": {
      "External id": 537878,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 3349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622635395.236, "dur": 5.292,
    "args": {
      "External id": 537879,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622635402.356, "dur": 13863.658,
    "args": {
      "External id": 537880,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 3351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622649426.036, "dur": 11.198,
    "args": {
      "External id": 537881,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 3352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622649429.371, "dur": 7.499,
    "args": {
      "External id": 537882,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622649469.249, "dur": 329.960,
    "args": {
      "External id": 537883,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 3354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622649493.528, "dur": 300.390,
    "args": {
      "External id": 537884,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3355, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622649506.037, "dur": 282.280,
    "args": {
      "External id": 537885,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 3356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622649819.566, "dur": 2.633,
    "args": {
      "External id": 537886,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3357, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622649883.886, "dur": 6.550,
    "args": {
      "External id": 537887,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622649933.817, "dur": 1.510,
    "args": {
      "External id": 537888,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622649951.117, "dur": 3.193,
    "args": {
      "External id": 537889,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622649965.185, "dur": 0.710,
    "args": {
      "External id": 537890,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622649975.560, "dur": 1.096,
    "args": {
      "External id": 537891,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622649986.364, "dur": 0.743,
    "args": {
      "External id": 537892,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622649996.388, "dur": 2.873,
    "args": {
      "External id": 537893,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650010.001, "dur": 2.017,
    "args": {
      "External id": 537894,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650020.867, "dur": 0.762,
    "args": {
      "External id": 537895,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622650132.547, "dur": 3345.142,
    "args": {
      "External id": 537896,"Record function id": 0, "Ev Idx": 3367
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.7)", "pid": 4183438, "tid": 31367,
    "ts": 679622650154.096, "dur": 1017.512,
    "args": {
      "External id": 537897,"Record function id": 0, "Ev Idx": 3368
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.7)", "pid": 4183438, "tid": 31367,
    "ts": 679622650170.805, "dur": 307.019,
    "args": {
      "External id": 537898,"Record function id": 0, "Ev Idx": 3369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650251.742, "dur": 4.682,
    "args": {
      "External id": 537899,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650259.607, "dur": 0.919,
    "args": {
      "External id": 537900,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650262.280, "dur": 2.931,
    "args": {
      "External id": 537901,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650266.641, "dur": 0.663,
    "args": {
      "External id": 537902,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650268.760, "dur": 0.905,
    "args": {
      "External id": 537903,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650271.109, "dur": 0.708,
    "args": {
      "External id": 537904,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650273.455, "dur": 1.498,
    "args": {
      "External id": 537905,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650276.231, "dur": 0.895,
    "args": {
      "External id": 537906,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650278.380, "dur": 0.618,
    "args": {
      "External id": 537907,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622650280.342, "dur": 0.635,
    "args": {
      "External id": 537908,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622650298.289, "dur": 150.726,
    "args": {
      "External id": 537909,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622650314.033, "dur": 130.578,
    "args": {
      "External id": 537910,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622650327.166, "dur": 15.251,
    "args": {
      "External id": 537911,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622650345.127, "dur": 66.874,
    "args": {
      "External id": 537912,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 3383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622650347.683, "dur": 64.012,
    "args": {
      "External id": 537913,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 3384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650351.410, "dur": 5.676,
    "args": {
      "External id": 537914,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622650362.334, "dur": 48.528,
    "args": {
      "External id": 537915,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 3386
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.6", "pid": 4183438, "tid": 31367,
    "ts": 679622650552.836, "dur": 611.185,
    "args": {
      "External id": 537916,"Record function id": 0, "Ev Idx": 3387
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.6)", "pid": 4183438, "tid": 31367,
    "ts": 679622650570.429, "dur": 577.284,
    "args": {
      "External id": 537917,"Record function id": 0, "Ev Idx": 3388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622650621.792, "dur": 4.820,
    "args": {
      "External id": 537918,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622650641.395, "dur": 67.493,
    "args": {
      "External id": 537919,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650645.737, "dur": 1.624,
    "args": {
      "External id": 537920,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650648.931, "dur": 0.663,
    "args": {
      "External id": 537921,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650650.873, "dur": 0.486,
    "args": {
      "External id": 537922,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650689.340, "dur": 0.730,
    "args": {
      "External id": 537923,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650695.429, "dur": 0.790,
    "args": {
      "External id": 537924,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650697.056, "dur": 2.368,
    "args": {
      "External id": 537925,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650700.663, "dur": 0.471,
    "args": {
      "External id": 537926,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650702.280, "dur": 0.619,
    "args": {
      "External id": 537927,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650704.300, "dur": 0.426,
    "args": {
      "External id": 537928,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622650720.002, "dur": 36.964,
    "args": {
      "External id": 537929,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622650791.733, "dur": 102.461,
    "args": {
      "External id": 537930,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 3401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622650802.045, "dur": 4.407,
    "args": {
      "External id": 537931,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622650814.996, "dur": 9.825,
    "args": {
      "External id": 537932,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622650819.268, "dur": 5.173,
    "args": {
      "External id": 537933,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 3404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650822.489, "dur": 0.663,
    "args": {
      "External id": 537934,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622650831.465, "dur": 22.822,
    "args": {
      "External id": 537935,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650833.265, "dur": 0.629,
    "args": {
      "External id": 537936,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650835.457, "dur": 0.646,
    "args": {
      "External id": 537937,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650837.607, "dur": 2.113,
    "args": {
      "External id": 537938,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650841.220, "dur": 0.445,
    "args": {
      "External id": 537939,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650842.849, "dur": 0.635,
    "args": {
      "External id": 537940,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650845.175, "dur": 0.335,
    "args": {
      "External id": 537941,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650846.849, "dur": 0.839,
    "args": {
      "External id": 537942,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650849.062, "dur": 0.348,
    "args": {
      "External id": 537943,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622650850.742, "dur": 0.475,
    "args": {
      "External id": 537944,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622650864.710, "dur": 22.458,
    "args": {
      "External id": 537945,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622650941.463, "dur": 118.396,
    "args": {
      "External id": 537946,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 3417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622650965.591, "dur": 90.618,
    "args": {
      "External id": 537947,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3418, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622650975.400, "dur": 76.997,
    "args": {
      "External id": 537948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 3419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622651073.524, "dur": 1.977,
    "args": {
      "External id": 537949,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3420, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622651179.484, "dur": 2273.906,
    "args": {
      "External id": 537950,"Sequence number": 6319432, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3421
    }
  },
  {
    "ph": "f", "id": 196, "pid": 4183438, "tid": 31367, "ts": 679622651179.484,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622651293.634, "dur": 103.914,
    "args": {
      "External id": 537951,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 3422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622651438.103, "dur": 36.624,
    "args": {
      "External id": 537952,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 3423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622651490.765, "dur": 42.020,
    "args": {
      "External id": 537953,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 3424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622651541.409, "dur": 26.375,
    "args": {
      "External id": 537954,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622651575.859, "dur": 33.864,
    "args": {
      "External id": 537955,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622651616.276, "dur": 21.690,
    "args": {
      "External id": 537956,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622651644.663, "dur": 92.204,
    "args": {
      "External id": 537957,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622651765.073, "dur": 27.561,
    "args": {
      "External id": 537958,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 3429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622651809.742, "dur": 29.860,
    "args": {
      "External id": 537959,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622651859.299, "dur": 25.012,
    "args": {
      "External id": 537960,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622651903.172, "dur": 21.434,
    "args": {
      "External id": 537961,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622651936.142, "dur": 64.315,
    "args": {
      "External id": 537962,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622652006.590, "dur": 40.857,
    "args": {
      "External id": 537963,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622652081.302, "dur": 198.352,
    "args": {
      "External id": 537964,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622652176.603, "dur": 8.623,
    "args": {
      "External id": 537965,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622652187.591, "dur": 3.355,
    "args": {
      "External id": 537966,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622652312.526, "dur": 28.755,
    "args": {
      "External id": 537967,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622652353.089, "dur": 15.996,
    "args": {
      "External id": 537968,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622652377.158, "dur": 52.540,
    "args": {
      "External id": 537969,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622652436.854, "dur": 42.424,
    "args": {
      "External id": 537970,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622652491.195, "dur": 40.992,
    "args": {
      "External id": 537971,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622652540.759, "dur": 38.898,
    "args": {
      "External id": 537972,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622652590.521, "dur": 58.055,
    "args": {
      "External id": 537973,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622652707.345, "dur": 57.097,
    "args": {
      "External id": 537974,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622652802.603, "dur": 445.843,
    "args": {
      "External id": 537975,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 3446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622653281.180, "dur": 36.216,
    "args": {
      "External id": 537976,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622653336.439, "dur": 18.352,
    "args": {
      "External id": 537977,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622653373.758, "dur": 14.811,
    "args": {
      "External id": 537978,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622653403.797, "dur": 16.064,
    "args": {
      "External id": 537979,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 3450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653508.812, "dur": 22.795,
    "args": {
      "External id": 537980,"Record function id": 0, "Ev Idx": 3451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653515.048, "dur": 15.146,
    "args": {
      "External id": 537981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653521.279, "dur": 7.904,
    "args": {
      "External id": 537982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653523.265, "dur": 5.736,
    "args": {
      "External id": 537983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653537.549, "dur": 6.967,
    "args": {
      "External id": 537984,"Record function id": 0, "Ev Idx": 3455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653539.709, "dur": 4.127,
    "args": {
      "External id": 537985,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653540.916, "dur": 2.278,
    "args": {
      "External id": 537986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653541.764, "dur": 1.289,
    "args": {
      "External id": 537987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653549.429, "dur": 6.172,
    "args": {
      "External id": 537988,"Record function id": 0, "Ev Idx": 3459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653551.238, "dur": 3.603,
    "args": {
      "External id": 537989,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653552.302, "dur": 1.854,
    "args": {
      "External id": 537990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653553.091, "dur": 0.891,
    "args": {
      "External id": 537991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653561.049, "dur": 7.083,
    "args": {
      "External id": 537992,"Record function id": 0, "Ev Idx": 3463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653563.742, "dur": 3.745,
    "args": {
      "External id": 537993,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653565.221, "dur": 1.676,
    "args": {
      "External id": 537994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653565.795, "dur": 0.934,
    "args": {
      "External id": 537995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653572.668, "dur": 8.913,
    "args": {
      "External id": 537996,"Record function id": 0, "Ev Idx": 3467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653574.738, "dur": 6.130,
    "args": {
      "External id": 537997,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653575.914, "dur": 4.274,
    "args": {
      "External id": 537998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653576.787, "dur": 3.227,
    "args": {
      "External id": 537999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653586.150, "dur": 5.939,
    "args": {
      "External id": 538000,"Record function id": 0, "Ev Idx": 3471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653587.982, "dur": 3.453,
    "args": {
      "External id": 538001,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653588.809, "dur": 2.040,
    "args": {
      "External id": 538002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653589.482, "dur": 1.203,
    "args": {
      "External id": 538003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653597.000, "dur": 6.588,
    "args": {
      "External id": 538004,"Record function id": 0, "Ev Idx": 3475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653599.113, "dur": 3.792,
    "args": {
      "External id": 538005,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653600.245, "dur": 2.091,
    "args": {
      "External id": 538006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653601.292, "dur": 0.894,
    "args": {
      "External id": 538007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653608.054, "dur": 4.775,
    "args": {
      "External id": 538008,"Record function id": 0, "Ev Idx": 3479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653609.670, "dur": 2.731,
    "args": {
      "External id": 538009,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653610.585, "dur": 1.326,
    "args": {
      "External id": 538010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653611.082, "dur": 0.764,
    "args": {
      "External id": 538011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653615.901, "dur": 3.895,
    "args": {
      "External id": 538012,"Record function id": 0, "Ev Idx": 3483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622653617.201, "dur": 2.218,
    "args": {
      "External id": 538013,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653617.655, "dur": 1.371,
    "args": {
      "External id": 538014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622653618.310, "dur": 0.647,
    "args": {
      "External id": 538015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622653623.249, "dur": 14754.449,
    "args": {
      "External id": 538016,"Record function id": 0, "Sequence number": 6319431, "Fwd thread id": 1, "Ev Idx": 3487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622653624.325, "dur": 14744.433,
    "args": {
      "External id": 538017,"Sequence number": 6319431, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3488
    }
  },
  {
    "ph": "f", "id": 197, "pid": 4183438, "tid": 31367, "ts": 679622653624.325,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.7)", "pid": 4183438, "tid": 31367,
    "ts": 679622653693.457, "dur": 48.507,
    "args": {
      "External id": 538018,"Record function id": 0, "Ev Idx": 3489
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.7)", "pid": 4183438, "tid": 31367,
    "ts": 679622653751.537, "dur": 69.263,
    "args": {
      "External id": 538019,"Record function id": 0, "Ev Idx": 3490
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.7)", "pid": 4183438, "tid": 31367,
    "ts": 679622653827.666, "dur": 14533.294,
    "args": {
      "External id": 538020,"Record function id": 0, "Ev Idx": 3491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622653924.963, "dur": 8.554,
    "args": {
      "External id": 538021,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622653943.816, "dur": 5.417,
    "args": {
      "External id": 538022,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622653963.049, "dur": 13601.618,
    "args": {
      "External id": 538023,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622653977.115, "dur": 13574.215,
    "args": {
      "External id": 538024,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622654001.298, "dur": 14.420,
    "args": {
      "External id": 538025,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622654019.483, "dur": 13494.900,
    "args": {
      "External id": 538026,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 3497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622654022.151, "dur": 13491.323,
    "args": {
      "External id": 538027,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 3498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622654026.587, "dur": 6.787,
    "args": {
      "External id": 538028,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622654035.151, "dur": 13474.204,
    "args": {
      "External id": 538029,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 3500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622667700.719, "dur": 15.423,
    "args": {
      "External id": 538030,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 3501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622667704.487, "dur": 10.544,
    "args": {
      "External id": 538031,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622667760.258, "dur": 289.864,
    "args": {
      "External id": 538032,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 3503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622667790.928, "dur": 254.511,
    "args": {
      "External id": 538033,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3504, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622667804.046, "dur": 236.611,
    "args": {
      "External id": 538034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 3505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622668067.812, "dur": 2.412,
    "args": {
      "External id": 538035,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3506, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668152.536, "dur": 7.034,
    "args": {
      "External id": 538036,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668203.638, "dur": 1.409,
    "args": {
      "External id": 538037,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668220.907, "dur": 1.137,
    "args": {
      "External id": 538038,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668232.753, "dur": 2.431,
    "args": {
      "External id": 538039,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668247.067, "dur": 0.639,
    "args": {
      "External id": 538040,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668257.730, "dur": 0.686,
    "args": {
      "External id": 538041,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668268.948, "dur": 0.643,
    "args": {
      "External id": 538042,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668279.785, "dur": 6.992,
    "args": {
      "External id": 538043,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668296.040, "dur": 0.854,
    "args": {
      "External id": 538044,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622668394.615, "dur": 3191.340,
    "args": {
      "External id": 538045,"Record function id": 0, "Ev Idx": 3516
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.6)", "pid": 4183438, "tid": 31367,
    "ts": 679622668414.418, "dur": 1024.756,
    "args": {
      "External id": 538046,"Record function id": 0, "Ev Idx": 3517
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.6)", "pid": 4183438, "tid": 31367,
    "ts": 679622668430.258, "dur": 346.144,
    "args": {
      "External id": 538047,"Record function id": 0, "Ev Idx": 3518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668507.764, "dur": 4.472,
    "args": {
      "External id": 538048,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668515.642, "dur": 0.800,
    "args": {
      "External id": 538049,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668518.704, "dur": 0.768,
    "args": {
      "External id": 538050,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668521.466, "dur": 0.517,
    "args": {
      "External id": 538051,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668523.584, "dur": 2.855,
    "args": {
      "External id": 538052,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668527.918, "dur": 0.799,
    "args": {
      "External id": 538053,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668530.399, "dur": 1.375,
    "args": {
      "External id": 538054,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668533.559, "dur": 0.770,
    "args": {
      "External id": 538055,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668536.004, "dur": 0.826,
    "args": {
      "External id": 538056,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622668538.415, "dur": 0.575,
    "args": {
      "External id": 538057,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622668557.017, "dur": 185.601,
    "args": {
      "External id": 538058,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622668573.133, "dur": 164.241,
    "args": {
      "External id": 538059,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622668587.083, "dur": 14.089,
    "args": {
      "External id": 538060,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622668604.055, "dur": 104.550,
    "args": {
      "External id": 538061,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 3532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622668606.468, "dur": 101.729,
    "args": {
      "External id": 538062,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 3533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668610.003, "dur": 5.822,
    "args": {
      "External id": 538063,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622668617.553, "dur": 89.570,
    "args": {
      "External id": 538064,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 3535
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.5", "pid": 4183438, "tid": 31367,
    "ts": 679622668855.479, "dur": 576.109,
    "args": {
      "External id": 538065,"Record function id": 0, "Ev Idx": 3536
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.5)", "pid": 4183438, "tid": 31367,
    "ts": 679622668873.221, "dur": 546.923,
    "args": {
      "External id": 538066,"Record function id": 0, "Ev Idx": 3537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622668937.578, "dur": 7.278,
    "args": {
      "External id": 538067,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622668959.544, "dur": 29.221,
    "args": {
      "External id": 538068,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668964.124, "dur": 5.290,
    "args": {
      "External id": 538069,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668971.248, "dur": 0.870,
    "args": {
      "External id": 538070,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668973.477, "dur": 0.504,
    "args": {
      "External id": 538071,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668975.200, "dur": 0.433,
    "args": {
      "External id": 538072,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668976.664, "dur": 0.316,
    "args": {
      "External id": 538073,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668978.420, "dur": 0.464,
    "args": {
      "External id": 538074,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668980.292, "dur": 0.284,
    "args": {
      "External id": 538075,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668981.449, "dur": 2.175,
    "args": {
      "External id": 538076,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622668984.640, "dur": 0.650,
    "args": {
      "External id": 538077,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622668998.916, "dur": 32.393,
    "args": {
      "External id": 538078,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622669060.976, "dur": 120.559,
    "args": {
      "External id": 538079,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 3550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622669070.567, "dur": 3.213,
    "args": {
      "External id": 538080,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622669079.980, "dur": 12.847,
    "args": {
      "External id": 538081,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622669087.149, "dur": 5.294,
    "args": {
      "External id": 538082,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 3553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669090.452, "dur": 0.825,
    "args": {
      "External id": 538083,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622669118.071, "dur": 22.365,
    "args": {
      "External id": 538084,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669120.709, "dur": 0.881,
    "args": {
      "External id": 538085,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669123.815, "dur": 0.362,
    "args": {
      "External id": 538086,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669125.479, "dur": 0.445,
    "args": {
      "External id": 538087,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669127.230, "dur": 0.419,
    "args": {
      "External id": 538088,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669128.835, "dur": 2.181,
    "args": {
      "External id": 538089,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669132.802, "dur": 0.380,
    "args": {
      "External id": 538090,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669134.324, "dur": 0.440,
    "args": {
      "External id": 538091,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669135.908, "dur": 0.614,
    "args": {
      "External id": 538092,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622669137.596, "dur": 0.272,
    "args": {
      "External id": 538093,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622669152.488, "dur": 21.611,
    "args": {
      "External id": 538094,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622669228.049, "dur": 120.493,
    "args": {
      "External id": 538095,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 3566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622669255.927, "dur": 88.872,
    "args": {
      "External id": 538096,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3567, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622669265.753, "dur": 75.143,
    "args": {
      "External id": 538097,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 3568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622669364.814, "dur": 1.741,
    "args": {
      "External id": 538098,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3569, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622669447.268, "dur": 2116.971,
    "args": {
      "External id": 538099,"Sequence number": 6319430, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3570
    }
  },
  {
    "ph": "f", "id": 198, "pid": 4183438, "tid": 31367, "ts": 679622669447.268,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622669560.280, "dur": 141.043,
    "args": {
      "External id": 538100,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 3571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622669745.742, "dur": 37.897,
    "args": {
      "External id": 538101,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 3572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622669800.944, "dur": 47.254,
    "args": {
      "External id": 538102,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 3573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622669857.959, "dur": 27.946,
    "args": {
      "External id": 538103,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622669892.195, "dur": 37.481,
    "args": {
      "External id": 538104,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622669938.547, "dur": 25.330,
    "args": {
      "External id": 538105,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622669970.859, "dur": 29.247,
    "args": {
      "External id": 538106,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622670022.555, "dur": 23.644,
    "args": {
      "External id": 538107,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 3578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622670063.224, "dur": 29.562,
    "args": {
      "External id": 538108,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622670127.810, "dur": 22.534,
    "args": {
      "External id": 538109,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622670164.123, "dur": 16.833,
    "args": {
      "External id": 538110,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670190.462, "dur": 41.779,
    "args": {
      "External id": 538111,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670235.980, "dur": 31.481,
    "args": {
      "External id": 538112,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622670311.787, "dur": 166.150,
    "args": {
      "External id": 538113,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622670387.543, "dur": 6.047,
    "args": {
      "External id": 538114,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622670395.347, "dur": 2.828,
    "args": {
      "External id": 538115,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622670506.332, "dur": 24.138,
    "args": {
      "External id": 538116,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622670542.245, "dur": 14.903,
    "args": {
      "External id": 538117,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670565.993, "dur": 46.335,
    "args": {
      "External id": 538118,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670620.044, "dur": 65.293,
    "args": {
      "External id": 538119,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670699.750, "dur": 44.623,
    "args": {
      "External id": 538120,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670753.345, "dur": 41.432,
    "args": {
      "External id": 538121,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670804.077, "dur": 37.537,
    "args": {
      "External id": 538122,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622670853.530, "dur": 66.769,
    "args": {
      "External id": 538123,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622670961.867, "dur": 36.950,
    "args": {
      "External id": 538124,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 3595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622671206.949, "dur": 33.810,
    "args": {
      "External id": 538125,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622671432.451, "dur": 23.634,
    "args": {
      "External id": 538126,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622671480.764, "dur": 18.835,
    "args": {
      "External id": 538127,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622671513.397, "dur": 17.837,
    "args": {
      "External id": 538128,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 3599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671610.601, "dur": 16.907,
    "args": {
      "External id": 538129,"Record function id": 0, "Ev Idx": 3600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671614.621, "dur": 12.097,
    "args": {
      "External id": 538130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671619.321, "dur": 6.624,
    "args": {
      "External id": 538131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671620.857, "dur": 5.002,
    "args": {
      "External id": 538132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671631.329, "dur": 5.200,
    "args": {
      "External id": 538133,"Record function id": 0, "Ev Idx": 3604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671632.740, "dur": 3.309,
    "args": {
      "External id": 538134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671633.829, "dur": 1.782,
    "args": {
      "External id": 538135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671634.610, "dur": 0.933,
    "args": {
      "External id": 538136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671639.761, "dur": 6.538,
    "args": {
      "External id": 538137,"Record function id": 0, "Ev Idx": 3608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671641.334, "dur": 4.439,
    "args": {
      "External id": 538138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671641.961, "dur": 3.385,
    "args": {
      "External id": 538139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671642.574, "dur": 2.673,
    "args": {
      "External id": 538140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671649.456, "dur": 4.440,
    "args": {
      "External id": 538141,"Record function id": 0, "Ev Idx": 3612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671650.852, "dur": 2.609,
    "args": {
      "External id": 538142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671651.627, "dur": 1.425,
    "args": {
      "External id": 538143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671652.107, "dur": 0.838,
    "args": {
      "External id": 538144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671690.125, "dur": 7.498,
    "args": {
      "External id": 538145,"Record function id": 0, "Ev Idx": 3616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671691.945, "dur": 4.996,
    "args": {
      "External id": 538146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671693.146, "dur": 2.848,
    "args": {
      "External id": 538147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671693.969, "dur": 1.834,
    "args": {
      "External id": 538148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671700.885, "dur": 4.280,
    "args": {
      "External id": 538149,"Record function id": 0, "Ev Idx": 3620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671702.356, "dur": 2.356,
    "args": {
      "External id": 538150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671702.892, "dur": 1.384,
    "args": {
      "External id": 538151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671703.369, "dur": 0.841,
    "args": {
      "External id": 538152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671708.519, "dur": 4.606,
    "args": {
      "External id": 538153,"Record function id": 0, "Ev Idx": 3624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671710.371, "dur": 2.354,
    "args": {
      "External id": 538154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671710.944, "dur": 1.310,
    "args": {
      "External id": 538155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671711.378, "dur": 0.811,
    "args": {
      "External id": 538156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671716.197, "dur": 3.758,
    "args": {
      "External id": 538157,"Record function id": 0, "Ev Idx": 3628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671717.407, "dur": 2.138,
    "args": {
      "External id": 538158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671717.924, "dur": 1.195,
    "args": {
      "External id": 538159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671718.352, "dur": 0.693,
    "args": {
      "External id": 538160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671723.001, "dur": 4.224,
    "args": {
      "External id": 538161,"Record function id": 0, "Ev Idx": 3632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622671724.281, "dur": 2.525,
    "args": {
      "External id": 538162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671724.956, "dur": 1.424,
    "args": {
      "External id": 538163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622671725.667, "dur": 0.642,
    "args": {
      "External id": 538164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622671731.168, "dur": 16328.335,
    "args": {
      "External id": 538165,"Record function id": 0, "Sequence number": 6319429, "Fwd thread id": 1, "Ev Idx": 3636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622671732.585, "dur": 16318.612,
    "args": {
      "External id": 538166,"Sequence number": 6319429, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3637
    }
  },
  {
    "ph": "f", "id": 199, "pid": 4183438, "tid": 31367, "ts": 679622671732.585,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.6)", "pid": 4183438, "tid": 31367,
    "ts": 679622671765.200, "dur": 41.123,
    "args": {
      "External id": 538167,"Record function id": 0, "Ev Idx": 3638
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.6)", "pid": 4183438, "tid": 31367,
    "ts": 679622671814.153, "dur": 63.500,
    "args": {
      "External id": 538168,"Record function id": 0, "Ev Idx": 3639
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.6)", "pid": 4183438, "tid": 31367,
    "ts": 679622671884.473, "dur": 16158.656,
    "args": {
      "External id": 538169,"Record function id": 0, "Ev Idx": 3640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622671977.918, "dur": 7.953,
    "args": {
      "External id": 538170,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622671995.695, "dur": 6.758,
    "args": {
      "External id": 538171,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622672016.265, "dur": 15257.072,
    "args": {
      "External id": 538172,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622672030.303, "dur": 15231.322,
    "args": {
      "External id": 538173,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622672056.653, "dur": 15.295,
    "args": {
      "External id": 538174,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622672076.031, "dur": 15139.834,
    "args": {
      "External id": 538175,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 3646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622672078.510, "dur": 15136.340,
    "args": {
      "External id": 538176,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 3647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622672082.890, "dur": 4.625,
    "args": {
      "External id": 538177,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622672089.159, "dur": 15120.434,
    "args": {
      "External id": 538178,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 3649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622687385.030, "dur": 11.946,
    "args": {
      "External id": 538179,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 3650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622687388.691, "dur": 7.844,
    "args": {
      "External id": 538180,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622687426.297, "dur": 328.442,
    "args": {
      "External id": 538181,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 3652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622687451.877, "dur": 297.446,
    "args": {
      "External id": 538182,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3653, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622687463.577, "dur": 279.927,
    "args": {
      "External id": 538183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 3654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622687774.043, "dur": 2.431,
    "args": {
      "External id": 538184,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3655, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687836.843, "dur": 7.058,
    "args": {
      "External id": 538185,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687890.588, "dur": 1.603,
    "args": {
      "External id": 538186,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687907.357, "dur": 4.730,
    "args": {
      "External id": 538187,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687923.349, "dur": 2.547,
    "args": {
      "External id": 538188,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687936.316, "dur": 1.135,
    "args": {
      "External id": 538189,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687946.811, "dur": 0.947,
    "args": {
      "External id": 538190,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687957.948, "dur": 0.981,
    "args": {
      "External id": 538191,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687969.582, "dur": 3.681,
    "args": {
      "External id": 538192,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622687983.126, "dur": 0.872,
    "args": {
      "External id": 538193,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622688074.799, "dur": 3108.778,
    "args": {
      "External id": 538194,"Record function id": 0, "Ev Idx": 3665
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.5)", "pid": 4183438, "tid": 31367,
    "ts": 679622688093.148, "dur": 1047.646,
    "args": {
      "External id": 538195,"Record function id": 0, "Ev Idx": 3666
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.5)", "pid": 4183438, "tid": 31367,
    "ts": 679622688128.415, "dur": 314.160,
    "args": {
      "External id": 538196,"Record function id": 0, "Ev Idx": 3667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688215.782, "dur": 4.786,
    "args": {
      "External id": 538197,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688224.190, "dur": 1.040,
    "args": {
      "External id": 538198,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688227.469, "dur": 0.774,
    "args": {
      "External id": 538199,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688229.828, "dur": 0.964,
    "args": {
      "External id": 538200,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688232.413, "dur": 2.563,
    "args": {
      "External id": 538201,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688236.811, "dur": 1.077,
    "args": {
      "External id": 538202,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688239.219, "dur": 1.187,
    "args": {
      "External id": 538203,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688241.723, "dur": 0.758,
    "args": {
      "External id": 538204,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688244.088, "dur": 0.698,
    "args": {
      "External id": 538205,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622688246.795, "dur": 0.815,
    "args": {
      "External id": 538206,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622688264.882, "dur": 145.257,
    "args": {
      "External id": 538207,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622688281.162, "dur": 124.681,
    "args": {
      "External id": 538208,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622688294.514, "dur": 13.806,
    "args": {
      "External id": 538209,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622688311.554, "dur": 65.073,
    "args": {
      "External id": 538210,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 3681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622688313.979, "dur": 62.328,
    "args": {
      "External id": 538211,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 3682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688317.971, "dur": 6.437,
    "args": {
      "External id": 538212,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622688326.107, "dur": 49.443,
    "args": {
      "External id": 538213,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 3684
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.4", "pid": 4183438, "tid": 31367,
    "ts": 679622688522.086, "dur": 611.344,
    "args": {
      "External id": 538214,"Record function id": 0, "Ev Idx": 3685
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.4)", "pid": 4183438, "tid": 31367,
    "ts": 679622688537.080, "dur": 581.029,
    "args": {
      "External id": 538215,"Record function id": 0, "Ev Idx": 3686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622688593.700, "dur": 6.510,
    "args": {
      "External id": 538216,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622688615.180, "dur": 25.949,
    "args": {
      "External id": 538217,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688619.485, "dur": 1.684,
    "args": {
      "External id": 538218,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688622.894, "dur": 0.524,
    "args": {
      "External id": 538219,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688625.127, "dur": 0.589,
    "args": {
      "External id": 538220,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688627.320, "dur": 0.427,
    "args": {
      "External id": 538221,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688628.909, "dur": 0.359,
    "args": {
      "External id": 538222,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688630.511, "dur": 0.436,
    "args": {
      "External id": 538223,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688632.151, "dur": 0.249,
    "args": {
      "External id": 538224,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688634.113, "dur": 2.460,
    "args": {
      "External id": 538225,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688637.906, "dur": 0.420,
    "args": {
      "External id": 538226,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622688690.924, "dur": 40.173,
    "args": {
      "External id": 538227,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622688768.924, "dur": 102.080,
    "args": {
      "External id": 538228,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 3699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622688779.687, "dur": 4.497,
    "args": {
      "External id": 538229,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622688789.689, "dur": 9.954,
    "args": {
      "External id": 538230,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622688793.675, "dur": 5.576,
    "args": {
      "External id": 538231,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 3702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688797.035, "dur": 0.638,
    "args": {
      "External id": 538232,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622688806.653, "dur": 25.791,
    "args": {
      "External id": 538233,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688808.890, "dur": 0.671,
    "args": {
      "External id": 538234,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688810.804, "dur": 0.533,
    "args": {
      "External id": 538235,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688812.570, "dur": 0.714,
    "args": {
      "External id": 538236,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688814.458, "dur": 0.616,
    "args": {
      "External id": 538237,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688816.400, "dur": 5.286,
    "args": {
      "External id": 538238,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688822.939, "dur": 0.195,
    "args": {
      "External id": 538239,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688824.230, "dur": 0.544,
    "args": {
      "External id": 538240,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688826.179, "dur": 0.551,
    "args": {
      "External id": 538241,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622688828.212, "dur": 0.575,
    "args": {
      "External id": 538242,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622688843.332, "dur": 20.212,
    "args": {
      "External id": 538243,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622688918.307, "dur": 121.387,
    "args": {
      "External id": 538244,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 3715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622688946.304, "dur": 89.792,
    "args": {
      "External id": 538245,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3716, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622688955.510, "dur": 75.732,
    "args": {
      "External id": 538246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 3717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622689053.342, "dur": 1.667,
    "args": {
      "External id": 538247,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3718, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622689148.486, "dur": 1832.105,
    "args": {
      "External id": 538248,"Sequence number": 6319428, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3719
    }
  },
  {
    "ph": "f", "id": 200, "pid": 4183438, "tid": 31367, "ts": 679622689148.486,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622689257.053, "dur": 101.681,
    "args": {
      "External id": 538249,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 3720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622689397.688, "dur": 36.180,
    "args": {
      "External id": 538250,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 3721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622689450.402, "dur": 42.804,
    "args": {
      "External id": 538251,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 3722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622689502.713, "dur": 28.636,
    "args": {
      "External id": 538252,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622689536.983, "dur": 32.909,
    "args": {
      "External id": 538253,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622689577.680, "dur": 20.886,
    "args": {
      "External id": 538254,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622689605.435, "dur": 28.459,
    "args": {
      "External id": 538255,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622689693.561, "dur": 26.489,
    "args": {
      "External id": 538256,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 3727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622689741.552, "dur": 27.835,
    "args": {
      "External id": 538257,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622689793.705, "dur": 18.403,
    "args": {
      "External id": 538258,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622689824.341, "dur": 15.808,
    "args": {
      "External id": 538259,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622689850.120, "dur": 41.108,
    "args": {
      "External id": 538260,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622689894.604, "dur": 31.647,
    "args": {
      "External id": 538261,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622689952.472, "dur": 177.950,
    "args": {
      "External id": 538262,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622690024.285, "dur": 6.323,
    "args": {
      "External id": 538263,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622690032.236, "dur": 2.878,
    "args": {
      "External id": 538264,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622690162.517, "dur": 25.828,
    "args": {
      "External id": 538265,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622690201.218, "dur": 14.668,
    "args": {
      "External id": 538266,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622690225.089, "dur": 48.577,
    "args": {
      "External id": 538267,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622690281.322, "dur": 31.964,
    "args": {
      "External id": 538268,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622690319.948, "dur": 28.189,
    "args": {
      "External id": 538269,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622690352.933, "dur": 29.130,
    "args": {
      "External id": 538270,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622690388.261, "dur": 26.447,
    "args": {
      "External id": 538271,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622690423.211, "dur": 38.497,
    "args": {
      "External id": 538272,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622690485.331, "dur": 31.633,
    "args": {
      "External id": 538273,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 3744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622690542.002, "dur": 32.355,
    "args": {
      "External id": 538274,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622690600.778, "dur": 37.791,
    "args": {
      "External id": 538275,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622690706.174, "dur": 24.268,
    "args": {
      "External id": 538276,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622690751.540, "dur": 23.299,
    "args": {
      "External id": 538277,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 3748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691212.117, "dur": 20.226,
    "args": {
      "External id": 538278,"Record function id": 0, "Ev Idx": 3749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691216.296, "dur": 14.712,
    "args": {
      "External id": 538279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691222.242, "dur": 7.487,
    "args": {
      "External id": 538280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691224.258, "dur": 5.301,
    "args": {
      "External id": 538281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691238.123, "dur": 6.505,
    "args": {
      "External id": 538282,"Record function id": 0, "Ev Idx": 3753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691240.363, "dur": 3.710,
    "args": {
      "External id": 538283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691241.292, "dur": 2.220,
    "args": {
      "External id": 538284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691242.115, "dur": 1.314,
    "args": {
      "External id": 538285,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691248.016, "dur": 4.848,
    "args": {
      "External id": 538286,"Record function id": 0, "Ev Idx": 3757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691249.501, "dur": 2.963,
    "args": {
      "External id": 538287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691250.241, "dur": 1.804,
    "args": {
      "External id": 538288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691250.851, "dur": 1.109,
    "args": {
      "External id": 538289,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691256.049, "dur": 5.436,
    "args": {
      "External id": 538290,"Record function id": 0, "Ev Idx": 3761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691257.428, "dur": 3.677,
    "args": {
      "External id": 538291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691258.192, "dur": 2.521,
    "args": {
      "External id": 538292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691258.473, "dur": 2.114,
    "args": {
      "External id": 538293,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691264.593, "dur": 3.976,
    "args": {
      "External id": 538294,"Record function id": 0, "Ev Idx": 3765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691265.789, "dur": 2.355,
    "args": {
      "External id": 538295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691266.309, "dur": 1.432,
    "args": {
      "External id": 538296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691266.681, "dur": 0.973,
    "args": {
      "External id": 538297,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691271.654, "dur": 4.494,
    "args": {
      "External id": 538298,"Record function id": 0, "Ev Idx": 3769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691273.025, "dur": 2.740,
    "args": {
      "External id": 538299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691273.758, "dur": 1.624,
    "args": {
      "External id": 538300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691274.569, "dur": 0.740,
    "args": {
      "External id": 538301,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691279.249, "dur": 4.401,
    "args": {
      "External id": 538302,"Record function id": 0, "Ev Idx": 3773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691280.622, "dur": 2.613,
    "args": {
      "External id": 538303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691281.317, "dur": 1.543,
    "args": {
      "External id": 538304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691281.999, "dur": 0.794,
    "args": {
      "External id": 538305,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691286.558, "dur": 3.552,
    "args": {
      "External id": 538306,"Record function id": 0, "Ev Idx": 3777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691287.786, "dur": 1.950,
    "args": {
      "External id": 538307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691288.287, "dur": 1.076,
    "args": {
      "External id": 538308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691288.692, "dur": 0.607,
    "args": {
      "External id": 538309,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691292.974, "dur": 3.568,
    "args": {
      "External id": 538310,"Record function id": 0, "Ev Idx": 3781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622691294.274, "dur": 1.882,
    "args": {
      "External id": 538311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691294.734, "dur": 1.068,
    "args": {
      "External id": 538312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622691295.105, "dur": 0.631,
    "args": {
      "External id": 538313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622691300.580, "dur": 15043.339,
    "args": {
      "External id": 538314,"Record function id": 0, "Sequence number": 6319427, "Fwd thread id": 1, "Ev Idx": 3785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622691302.240, "dur": 15032.129,
    "args": {
      "External id": 538315,"Sequence number": 6319427, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3786
    }
  },
  {
    "ph": "f", "id": 201, "pid": 4183438, "tid": 31367, "ts": 679622691302.240,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.5)", "pid": 4183438, "tid": 31367,
    "ts": 679622691334.854, "dur": 41.756,
    "args": {
      "External id": 538316,"Record function id": 0, "Ev Idx": 3787
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.5)", "pid": 4183438, "tid": 31367,
    "ts": 679622691384.453, "dur": 67.882,
    "args": {
      "External id": 538317,"Record function id": 0, "Ev Idx": 3788
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.5)", "pid": 4183438, "tid": 31367,
    "ts": 679622691459.118, "dur": 14866.835,
    "args": {
      "External id": 538318,"Record function id": 0, "Ev Idx": 3789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622691551.707, "dur": 8.090,
    "args": {
      "External id": 538319,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622691569.472, "dur": 5.228,
    "args": {
      "External id": 538320,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622691588.180, "dur": 14011.081,
    "args": {
      "External id": 538321,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622691602.515, "dur": 13986.431,
    "args": {
      "External id": 538322,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622691623.759, "dur": 16.268,
    "args": {
      "External id": 538323,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622691644.040, "dur": 13908.086,
    "args": {
      "External id": 538324,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 3795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622691646.488, "dur": 13904.788,
    "args": {
      "External id": 538325,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 3796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622691685.531, "dur": 6.579,
    "args": {
      "External id": 538326,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622691694.417, "dur": 13852.980,
    "args": {
      "External id": 538327,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 3798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622705734.394, "dur": 10.936,
    "args": {
      "External id": 538328,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 3799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622705737.615, "dur": 7.211,
    "args": {
      "External id": 538329,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622705776.105, "dur": 258.336,
    "args": {
      "External id": 538330,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 3801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622705802.517, "dur": 227.242,
    "args": {
      "External id": 538331,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3802, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622705814.197, "dur": 210.504,
    "args": {
      "External id": 538332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 3803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622706049.980, "dur": 2.227,
    "args": {
      "External id": 538333,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3804, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706128.065, "dur": 6.877,
    "args": {
      "External id": 538334,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706179.586, "dur": 1.424,
    "args": {
      "External id": 538335,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706195.652, "dur": 1.450,
    "args": {
      "External id": 538336,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706208.304, "dur": 2.509,
    "args": {
      "External id": 538337,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706221.119, "dur": 0.844,
    "args": {
      "External id": 538338,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706231.406, "dur": 1.036,
    "args": {
      "External id": 538339,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706242.449, "dur": 1.092,
    "args": {
      "External id": 538340,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706254.195, "dur": 2.349,
    "args": {
      "External id": 538341,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706265.317, "dur": 0.834,
    "args": {
      "External id": 538342,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622706359.766, "dur": 3144.630,
    "args": {
      "External id": 538343,"Record function id": 0, "Ev Idx": 3814
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.4)", "pid": 4183438, "tid": 31367,
    "ts": 679622706378.647, "dur": 1027.215,
    "args": {
      "External id": 538344,"Record function id": 0, "Ev Idx": 3815
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.4)", "pid": 4183438, "tid": 31367,
    "ts": 679622706392.910, "dur": 341.427,
    "args": {
      "External id": 538345,"Record function id": 0, "Ev Idx": 3816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706469.436, "dur": 4.180,
    "args": {
      "External id": 538346,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706476.831, "dur": 1.217,
    "args": {
      "External id": 538347,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706480.016, "dur": 1.172,
    "args": {
      "External id": 538348,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706482.792, "dur": 0.673,
    "args": {
      "External id": 538349,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706484.812, "dur": 2.865,
    "args": {
      "External id": 538350,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706489.642, "dur": 1.003,
    "args": {
      "External id": 538351,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706492.367, "dur": 0.982,
    "args": {
      "External id": 538352,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706494.967, "dur": 1.001,
    "args": {
      "External id": 538353,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706497.554, "dur": 1.205,
    "args": {
      "External id": 538354,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622706500.223, "dur": 1.176,
    "args": {
      "External id": 538355,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622706518.771, "dur": 183.986,
    "args": {
      "External id": 538356,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622706534.720, "dur": 162.330,
    "args": {
      "External id": 538357,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622706546.702, "dur": 15.596,
    "args": {
      "External id": 538358,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622706565.176, "dur": 65.137,
    "args": {
      "External id": 538359,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 3830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622706567.552, "dur": 62.497,
    "args": {
      "External id": 538360,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 3831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706571.167, "dur": 5.240,
    "args": {
      "External id": 538361,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622706577.992, "dur": 51.442,
    "args": {
      "External id": 538362,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 3833
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.3", "pid": 4183438, "tid": 31367,
    "ts": 679622706822.391, "dur": 576.311,
    "args": {
      "External id": 538363,"Record function id": 0, "Ev Idx": 3834
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.3)", "pid": 4183438, "tid": 31367,
    "ts": 679622706839.621, "dur": 547.464,
    "args": {
      "External id": 538364,"Record function id": 0, "Ev Idx": 3835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622706900.606, "dur": 7.551,
    "args": {
      "External id": 538365,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622706924.108, "dur": 31.049,
    "args": {
      "External id": 538366,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706928.761, "dur": 1.987,
    "args": {
      "External id": 538367,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706935.655, "dur": 0.563,
    "args": {
      "External id": 538368,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706937.573, "dur": 0.282,
    "args": {
      "External id": 538369,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706939.452, "dur": 0.492,
    "args": {
      "External id": 538370,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706941.542, "dur": 0.477,
    "args": {
      "External id": 538371,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706943.473, "dur": 0.673,
    "args": {
      "External id": 538372,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706945.348, "dur": 0.270,
    "args": {
      "External id": 538373,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706947.171, "dur": 2.366,
    "args": {
      "External id": 538374,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622706950.694, "dur": 0.358,
    "args": {
      "External id": 538375,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622706964.432, "dur": 37.118,
    "args": {
      "External id": 538376,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622707031.723, "dur": 125.936,
    "args": {
      "External id": 538377,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 3848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622707041.149, "dur": 3.496,
    "args": {
      "External id": 538378,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622707053.405, "dur": 9.991,
    "args": {
      "External id": 538379,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622707057.475, "dur": 5.514,
    "args": {
      "External id": 538380,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 3851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707061.017, "dur": 0.681,
    "args": {
      "External id": 538381,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622707069.626, "dur": 25.701,
    "args": {
      "External id": 538382,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707071.463, "dur": 0.561,
    "args": {
      "External id": 538383,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707074.012, "dur": 0.414,
    "args": {
      "External id": 538384,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707076.035, "dur": 0.641,
    "args": {
      "External id": 538385,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707078.278, "dur": 0.512,
    "args": {
      "External id": 538386,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707080.335, "dur": 2.810,
    "args": {
      "External id": 538387,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707084.540, "dur": 0.341,
    "args": {
      "External id": 538388,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707086.268, "dur": 0.316,
    "args": {
      "External id": 538389,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707088.275, "dur": 0.645,
    "args": {
      "External id": 538390,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622707090.959, "dur": 0.426,
    "args": {
      "External id": 538391,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622707126.226, "dur": 22.726,
    "args": {
      "External id": 538392,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622707207.565, "dur": 116.980,
    "args": {
      "External id": 538393,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 3864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622707231.471, "dur": 86.411,
    "args": {
      "External id": 538394,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3865, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622707240.929, "dur": 72.540,
    "args": {
      "External id": 538395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 3866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622707337.486, "dur": 1.819,
    "args": {
      "External id": 538396,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3867, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622707412.431, "dur": 2071.207,
    "args": {
      "External id": 538397,"Sequence number": 6319426, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3868
    }
  },
  {
    "ph": "f", "id": 202, "pid": 4183438, "tid": 31367, "ts": 679622707412.431,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622707521.170, "dur": 98.864,
    "args": {
      "External id": 538398,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 3869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622707699.704, "dur": 38.931,
    "args": {
      "External id": 538399,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 3870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622707760.657, "dur": 51.734,
    "args": {
      "External id": 538400,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 3871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622707822.297, "dur": 26.297,
    "args": {
      "External id": 538401,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622707854.443, "dur": 32.685,
    "args": {
      "External id": 538402,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622707895.214, "dur": 21.381,
    "args": {
      "External id": 538403,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 3874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622707923.421, "dur": 28.968,
    "args": {
      "External id": 538404,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 3875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622707975.578, "dur": 21.435,
    "args": {
      "External id": 538405,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 3876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622708014.448, "dur": 28.208,
    "args": {
      "External id": 538406,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622708061.700, "dur": 18.017,
    "args": {
      "External id": 538407,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622708091.620, "dur": 29.200,
    "args": {
      "External id": 538408,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708133.339, "dur": 41.155,
    "args": {
      "External id": 538409,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708178.435, "dur": 31.643,
    "args": {
      "External id": 538410,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622708236.355, "dur": 172.369,
    "args": {
      "External id": 538411,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622708311.764, "dur": 5.847,
    "args": {
      "External id": 538412,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622708319.603, "dur": 2.799,
    "args": {
      "External id": 538413,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622708439.577, "dur": 24.919,
    "args": {
      "External id": 538414,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622708476.533, "dur": 15.985,
    "args": {
      "External id": 538415,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708500.632, "dur": 44.875,
    "args": {
      "External id": 538416,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708552.791, "dur": 31.345,
    "args": {
      "External id": 538417,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708590.419, "dur": 27.439,
    "args": {
      "External id": 538418,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708622.308, "dur": 71.628,
    "args": {
      "External id": 538419,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708709.646, "dur": 49.505,
    "args": {
      "External id": 538420,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 3891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622708771.980, "dur": 41.417,
    "args": {
      "External id": 538421,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 3892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622708857.006, "dur": 42.384,
    "args": {
      "External id": 538422,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 3893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622708927.483, "dur": 33.866,
    "args": {
      "External id": 538423,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622709161.823, "dur": 27.114,
    "args": {
      "External id": 538424,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 3895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622709384.887, "dur": 22.951,
    "args": {
      "External id": 538425,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 3896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622709426.137, "dur": 23.964,
    "args": {
      "External id": 538426,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 3897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709528.289, "dur": 16.465,
    "args": {
      "External id": 538427,"Record function id": 0, "Ev Idx": 3898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709531.451, "dur": 12.331,
    "args": {
      "External id": 538428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709536.490, "dur": 6.398,
    "args": {
      "External id": 538429,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709538.229, "dur": 4.546,
    "args": {
      "External id": 538430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709548.579, "dur": 5.325,
    "args": {
      "External id": 538431,"Record function id": 0, "Ev Idx": 3902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709549.942, "dur": 3.540,
    "args": {
      "External id": 538432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709550.926, "dur": 2.126,
    "args": {
      "External id": 538433,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709551.602, "dur": 1.324,
    "args": {
      "External id": 538434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709557.202, "dur": 4.603,
    "args": {
      "External id": 538435,"Record function id": 0, "Ev Idx": 3906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709558.627, "dur": 2.729,
    "args": {
      "External id": 538436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709559.373, "dur": 1.559,
    "args": {
      "External id": 538437,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709559.964, "dur": 0.886,
    "args": {
      "External id": 538438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 3909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709565.158, "dur": 5.776,
    "args": {
      "External id": 538439,"Record function id": 0, "Ev Idx": 3910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709566.412, "dur": 4.117,
    "args": {
      "External id": 538440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709567.120, "dur": 2.975,
    "args": {
      "External id": 538441,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709567.540, "dur": 2.434,
    "args": {
      "External id": 538442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 3913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709574.117, "dur": 4.024,
    "args": {
      "External id": 538443,"Record function id": 0, "Ev Idx": 3914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709575.461, "dur": 2.252,
    "args": {
      "External id": 538444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709575.979, "dur": 1.332,
    "args": {
      "External id": 538445,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709576.438, "dur": 0.762,
    "args": {
      "External id": 538446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709581.141, "dur": 3.661,
    "args": {
      "External id": 538447,"Record function id": 0, "Ev Idx": 3918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709582.261, "dur": 2.138,
    "args": {
      "External id": 538448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709582.862, "dur": 1.123,
    "args": {
      "External id": 538449,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709583.300, "dur": 0.593,
    "args": {
      "External id": 538450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709587.984, "dur": 3.731,
    "args": {
      "External id": 538451,"Record function id": 0, "Ev Idx": 3922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709589.148, "dur": 2.145,
    "args": {
      "External id": 538452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709589.687, "dur": 1.172,
    "args": {
      "External id": 538453,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709590.037, "dur": 0.722,
    "args": {
      "External id": 538454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709594.671, "dur": 4.438,
    "args": {
      "External id": 538455,"Record function id": 0, "Ev Idx": 3926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709596.000, "dur": 2.723,
    "args": {
      "External id": 538456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709596.568, "dur": 1.726,
    "args": {
      "External id": 538457,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709597.315, "dur": 0.882,
    "args": {
      "External id": 538458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 3929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709602.015, "dur": 6.223,
    "args": {
      "External id": 538459,"Record function id": 0, "Ev Idx": 3930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622709603.280, "dur": 4.545,
    "args": {
      "External id": 538460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709603.854, "dur": 3.552,
    "args": {
      "External id": 538461,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622709606.533, "dur": 0.780,
    "args": {
      "External id": 538462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 3933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622709612.046, "dur": 14915.323,
    "args": {
      "External id": 538463,"Record function id": 0, "Sequence number": 6319425, "Fwd thread id": 1, "Ev Idx": 3934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622709613.365, "dur": 14905.580,
    "args": {
      "External id": 538464,"Sequence number": 6319425, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 3935
    }
  },
  {
    "ph": "f", "id": 203, "pid": 4183438, "tid": 31367, "ts": 679622709613.365,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.4)", "pid": 4183438, "tid": 31367,
    "ts": 679622709644.999, "dur": 73.574,
    "args": {
      "External id": 538465,"Record function id": 0, "Ev Idx": 3936
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.4)", "pid": 4183438, "tid": 31367,
    "ts": 679622709728.128, "dur": 65.203,
    "args": {
      "External id": 538466,"Record function id": 0, "Ev Idx": 3937
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.4)", "pid": 4183438, "tid": 31367,
    "ts": 679622709799.497, "dur": 14711.124,
    "args": {
      "External id": 538467,"Record function id": 0, "Ev Idx": 3938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622709892.039, "dur": 7.987,
    "args": {
      "External id": 538468,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622709910.677, "dur": 5.360,
    "args": {
      "External id": 538469,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622709929.594, "dur": 13884.116,
    "args": {
      "External id": 538470,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622709943.076, "dur": 13860.168,
    "args": {
      "External id": 538471,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 3942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622709965.387, "dur": 17.980,
    "args": {
      "External id": 538472,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622709987.630, "dur": 13772.164,
    "args": {
      "External id": 538473,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 3944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622709990.479, "dur": 13768.404,
    "args": {
      "External id": 538474,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 3945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622709994.814, "dur": 5.313,
    "args": {
      "External id": 538475,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622710001.678, "dur": 13753.012,
    "args": {
      "External id": 538476,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 3947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622723913.128, "dur": 11.642,
    "args": {
      "External id": 538477,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 3948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622723916.632, "dur": 7.677,
    "args": {
      "External id": 538478,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622723955.591, "dur": 272.124,
    "args": {
      "External id": 538479,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 3950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622723982.276, "dur": 240.554,
    "args": {
      "External id": 538480,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3951, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622723993.912, "dur": 223.315,
    "args": {
      "External id": 538481,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 3952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622724245.880, "dur": 2.457,
    "args": {
      "External id": 538482,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3953, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724304.586, "dur": 6.837,
    "args": {
      "External id": 538483,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724358.704, "dur": 1.563,
    "args": {
      "External id": 538484,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724375.952, "dur": 1.240,
    "args": {
      "External id": 538485,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724389.859, "dur": 2.280,
    "args": {
      "External id": 538486,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724403.979, "dur": 1.253,
    "args": {
      "External id": 538487,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724415.091, "dur": 1.080,
    "args": {
      "External id": 538488,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724425.761, "dur": 0.958,
    "args": {
      "External id": 538489,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724437.088, "dur": 3.001,
    "args": {
      "External id": 538490,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724450.512, "dur": 1.295,
    "args": {
      "External id": 538491,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622724541.431, "dur": 3223.414,
    "args": {
      "External id": 538492,"Record function id": 0, "Ev Idx": 3963
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.3)", "pid": 4183438, "tid": 31367,
    "ts": 679622724561.098, "dur": 1059.120,
    "args": {
      "External id": 538493,"Record function id": 0, "Ev Idx": 3964
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.3)", "pid": 4183438, "tid": 31367,
    "ts": 679622724576.873, "dur": 384.165,
    "args": {
      "External id": 538494,"Record function id": 0, "Ev Idx": 3965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724708.590, "dur": 5.607,
    "args": {
      "External id": 538495,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 3966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724719.308, "dur": 1.077,
    "args": {
      "External id": 538496,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724722.267, "dur": 1.303,
    "args": {
      "External id": 538497,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724725.374, "dur": 1.086,
    "args": {
      "External id": 538498,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724737.582, "dur": 2.787,
    "args": {
      "External id": 538499,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724742.070, "dur": 1.093,
    "args": {
      "External id": 538500,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 3971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724744.876, "dur": 0.938,
    "args": {
      "External id": 538501,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 3972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724747.354, "dur": 1.232,
    "args": {
      "External id": 538502,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724750.410, "dur": 1.040,
    "args": {
      "External id": 538503,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622724753.130, "dur": 1.033,
    "args": {
      "External id": 538504,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 3975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622724773.425, "dur": 153.244,
    "args": {
      "External id": 538505,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622724794.454, "dur": 127.647,
    "args": {
      "External id": 538506,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 3977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622724807.631, "dur": 15.007,
    "args": {
      "External id": 538507,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622724825.551, "dur": 68.052,
    "args": {
      "External id": 538508,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 3979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622724828.434, "dur": 64.839,
    "args": {
      "External id": 538509,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 3980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622724832.603, "dur": 5.766,
    "args": {
      "External id": 538510,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622724840.147, "dur": 52.415,
    "args": {
      "External id": 538511,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 3982
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.2", "pid": 4183438, "tid": 31367,
    "ts": 679622725047.706, "dur": 564.311,
    "args": {
      "External id": 538512,"Record function id": 0, "Ev Idx": 3983
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.2)", "pid": 4183438, "tid": 31367,
    "ts": 679622725068.055, "dur": 532.301,
    "args": {
      "External id": 538513,"Record function id": 0, "Ev Idx": 3984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622725143.825, "dur": 7.533,
    "args": {
      "External id": 538514,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622725166.946, "dur": 30.366,
    "args": {
      "External id": 538515,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 3986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725174.946, "dur": 1.847,
    "args": {
      "External id": 538516,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725178.497, "dur": 0.287,
    "args": {
      "External id": 538517,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725179.959, "dur": 0.667,
    "args": {
      "External id": 538518,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725182.050, "dur": 0.816,
    "args": {
      "External id": 538519,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725183.886, "dur": 0.584,
    "args": {
      "External id": 538520,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725185.977, "dur": 0.561,
    "args": {
      "External id": 538521,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725188.130, "dur": 0.509,
    "args": {
      "External id": 538522,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725189.831, "dur": 2.006,
    "args": {
      "External id": 538523,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725192.790, "dur": 0.480,
    "args": {
      "External id": 538524,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 3995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622725207.198, "dur": 36.886,
    "args": {
      "External id": 538525,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 3996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622725274.449, "dur": 99.445,
    "args": {
      "External id": 538526,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 3997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622725284.320, "dur": 3.550,
    "args": {
      "External id": 538527,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622725292.811, "dur": 14.139,
    "args": {
      "External id": 538528,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 3999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622725301.207, "dur": 5.347,
    "args": {
      "External id": 538529,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 4000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725304.641, "dur": 0.642,
    "args": {
      "External id": 538530,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622725313.411, "dur": 21.490,
    "args": {
      "External id": 538531,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725315.630, "dur": 0.835,
    "args": {
      "External id": 538532,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725318.015, "dur": 0.501,
    "args": {
      "External id": 538533,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725319.722, "dur": 0.535,
    "args": {
      "External id": 538534,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725321.422, "dur": 0.644,
    "args": {
      "External id": 538535,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725323.262, "dur": 1.716,
    "args": {
      "External id": 538536,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725326.240, "dur": 0.461,
    "args": {
      "External id": 538537,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725328.072, "dur": 0.191,
    "args": {
      "External id": 538538,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725329.561, "dur": 0.368,
    "args": {
      "External id": 538539,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622725330.865, "dur": 0.357,
    "args": {
      "External id": 538540,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622725344.038, "dur": 22.517,
    "args": {
      "External id": 538541,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622725420.426, "dur": 115.999,
    "args": {
      "External id": 538542,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 4013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622725444.239, "dur": 88.515,
    "args": {
      "External id": 538543,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4014, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622725453.759, "dur": 74.818,
    "args": {
      "External id": 538544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 4015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622725550.264, "dur": 2.019,
    "args": {
      "External id": 538545,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4016, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622725627.288, "dur": 2115.068,
    "args": {
      "External id": 538546,"Sequence number": 6319424, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4017
    }
  },
  {
    "ph": "f", "id": 204, "pid": 4183438, "tid": 31367, "ts": 679622725627.288,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622725784.705, "dur": 108.410,
    "args": {
      "External id": 538547,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622725933.901, "dur": 39.571,
    "args": {
      "External id": 538548,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 4019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622725989.880, "dur": 42.162,
    "args": {
      "External id": 538549,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 4020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726041.180, "dur": 27.438,
    "args": {
      "External id": 538550,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726074.211, "dur": 51.338,
    "args": {
      "External id": 538551,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726137.316, "dur": 25.949,
    "args": {
      "External id": 538552,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726170.768, "dur": 30.537,
    "args": {
      "External id": 538553,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622726224.359, "dur": 24.568,
    "args": {
      "External id": 538554,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 4025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622726266.169, "dur": 27.931,
    "args": {
      "External id": 538555,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622726312.451, "dur": 17.231,
    "args": {
      "External id": 538556,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622726340.998, "dur": 15.177,
    "args": {
      "External id": 538557,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726365.895, "dur": 35.694,
    "args": {
      "External id": 538558,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726405.184, "dur": 29.542,
    "args": {
      "External id": 538559,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622726460.302, "dur": 164.039,
    "args": {
      "External id": 538560,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622726534.759, "dur": 5.989,
    "args": {
      "External id": 538561,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622726542.575, "dur": 3.019,
    "args": {
      "External id": 538562,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622726691.431, "dur": 29.030,
    "args": {
      "External id": 538563,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622726733.761, "dur": 15.864,
    "args": {
      "External id": 538564,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726759.761, "dur": 54.972,
    "args": {
      "External id": 538565,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726822.236, "dur": 34.184,
    "args": {
      "External id": 538566,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726863.055, "dur": 33.926,
    "args": {
      "External id": 538567,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726903.998, "dur": 41.611,
    "args": {
      "External id": 538568,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622726955.411, "dur": 40.151,
    "args": {
      "External id": 538569,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622727006.834, "dur": 54.573,
    "args": {
      "External id": 538570,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622727115.481, "dur": 43.990,
    "args": {
      "External id": 538571,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 4042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622727188.799, "dur": 200.211,
    "args": {
      "External id": 538572,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622727407.503, "dur": 190.499,
    "args": {
      "External id": 538573,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622727618.982, "dur": 18.636,
    "args": {
      "External id": 538574,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622727653.794, "dur": 52.483,
    "args": {
      "External id": 538575,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 4046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727789.009, "dur": 16.243,
    "args": {
      "External id": 538576,"Record function id": 0, "Ev Idx": 4047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727792.312, "dur": 12.048,
    "args": {
      "External id": 538577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727797.132, "dur": 6.273,
    "args": {
      "External id": 538578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727798.882, "dur": 4.415,
    "args": {
      "External id": 538579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727808.730, "dur": 4.647,
    "args": {
      "External id": 538580,"Record function id": 0, "Ev Idx": 4051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727810.087, "dur": 2.839,
    "args": {
      "External id": 538581,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727810.826, "dur": 1.609,
    "args": {
      "External id": 538582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727811.517, "dur": 0.838,
    "args": {
      "External id": 538583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727816.693, "dur": 4.940,
    "args": {
      "External id": 538584,"Record function id": 0, "Ev Idx": 4055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727818.061, "dur": 3.146,
    "args": {
      "External id": 538585,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727818.893, "dur": 1.863,
    "args": {
      "External id": 538586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727819.461, "dur": 1.215,
    "args": {
      "External id": 538587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727824.708, "dur": 5.927,
    "args": {
      "External id": 538588,"Record function id": 0, "Ev Idx": 4059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727826.208, "dur": 3.982,
    "args": {
      "External id": 538589,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727826.902, "dur": 2.866,
    "args": {
      "External id": 538590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727827.404, "dur": 2.306,
    "args": {
      "External id": 538591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727833.754, "dur": 4.207,
    "args": {
      "External id": 538592,"Record function id": 0, "Ev Idx": 4063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727834.954, "dur": 2.562,
    "args": {
      "External id": 538593,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727835.601, "dur": 1.488,
    "args": {
      "External id": 538594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727836.398, "dur": 0.632,
    "args": {
      "External id": 538595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727841.055, "dur": 4.703,
    "args": {
      "External id": 538596,"Record function id": 0, "Ev Idx": 4067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727842.368, "dur": 2.995,
    "args": {
      "External id": 538597,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727842.950, "dur": 1.974,
    "args": {
      "External id": 538598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727843.642, "dur": 1.216,
    "args": {
      "External id": 538599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727848.918, "dur": 4.232,
    "args": {
      "External id": 538600,"Record function id": 0, "Ev Idx": 4071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727850.316, "dur": 2.398,
    "args": {
      "External id": 538601,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727850.934, "dur": 1.264,
    "args": {
      "External id": 538602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727851.522, "dur": 0.611,
    "args": {
      "External id": 538603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727856.163, "dur": 4.289,
    "args": {
      "External id": 538604,"Record function id": 0, "Ev Idx": 4075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727857.441, "dur": 2.599,
    "args": {
      "External id": 538605,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727858.107, "dur": 1.502,
    "args": {
      "External id": 538606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727858.870, "dur": 0.675,
    "args": {
      "External id": 538607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727863.834, "dur": 3.745,
    "args": {
      "External id": 538608,"Record function id": 0, "Ev Idx": 4079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622727865.020, "dur": 2.162,
    "args": {
      "External id": 538609,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727865.657, "dur": 1.130,
    "args": {
      "External id": 538610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622727866.246, "dur": 0.474,
    "args": {
      "External id": 538611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622727871.200, "dur": 18166.128,
    "args": {
      "External id": 538612,"Record function id": 0, "Sequence number": 6319423, "Fwd thread id": 1, "Ev Idx": 4083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622727872.420, "dur": 18156.183,
    "args": {
      "External id": 538613,"Sequence number": 6319423, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4084
    }
  },
  {
    "ph": "f", "id": 205, "pid": 4183438, "tid": 31367, "ts": 679622727872.420,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.3)", "pid": 4183438, "tid": 31367,
    "ts": 679622727903.128, "dur": 40.886,
    "args": {
      "External id": 538614,"Record function id": 0, "Ev Idx": 4085
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.3)", "pid": 4183438, "tid": 31367,
    "ts": 679622727952.389, "dur": 61.861,
    "args": {
      "External id": 538615,"Record function id": 0, "Ev Idx": 4086
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.3)", "pid": 4183438, "tid": 31367,
    "ts": 679622728020.436, "dur": 18000.422,
    "args": {
      "External id": 538616,"Record function id": 0, "Ev Idx": 4087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622728128.027, "dur": 9.257,
    "args": {
      "External id": 538617,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622728148.414, "dur": 5.486,
    "args": {
      "External id": 538618,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622728168.640, "dur": 17069.000,
    "args": {
      "External id": 538619,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622728182.702, "dur": 17042.960,
    "args": {
      "External id": 538620,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622728202.900, "dur": 16.656,
    "args": {
      "External id": 538621,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622728224.016, "dur": 16955.518,
    "args": {
      "External id": 538622,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 4093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622728226.732, "dur": 16951.797,
    "args": {
      "External id": 538623,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 4094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622728231.006, "dur": 5.193,
    "args": {
      "External id": 538624,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622728249.383, "dur": 16924.073,
    "args": {
      "External id": 538625,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 4096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622745344.787, "dur": 12.246,
    "args": {
      "External id": 538626,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 4097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622745348.300, "dur": 8.374,
    "args": {
      "External id": 538627,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622745387.539, "dur": 337.764,
    "args": {
      "External id": 538628,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 4099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622745419.144, "dur": 301.029,
    "args": {
      "External id": 538629,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4100, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622745430.972, "dur": 283.292,
    "args": {
      "External id": 538630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 4101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622745744.262, "dur": 2.136,
    "args": {
      "External id": 538631,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4102, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745813.491, "dur": 7.247,
    "args": {
      "External id": 538632,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745865.521, "dur": 1.807,
    "args": {
      "External id": 538633,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745882.780, "dur": 1.211,
    "args": {
      "External id": 538634,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745895.591, "dur": 2.427,
    "args": {
      "External id": 538635,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745908.587, "dur": 0.986,
    "args": {
      "External id": 538636,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745919.715, "dur": 0.867,
    "args": {
      "External id": 538637,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745932.732, "dur": 0.932,
    "args": {
      "External id": 538638,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745944.878, "dur": 1.861,
    "args": {
      "External id": 538639,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622745956.452, "dur": 0.880,
    "args": {
      "External id": 538640,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622746054.399, "dur": 2711.104,
    "args": {
      "External id": 538641,"Record function id": 0, "Ev Idx": 4112
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.2)", "pid": 4183438, "tid": 31367,
    "ts": 679622746073.631, "dur": 1002.507,
    "args": {
      "External id": 538642,"Record function id": 0, "Ev Idx": 4113
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.2)", "pid": 4183438, "tid": 31367,
    "ts": 679622746089.442, "dur": 318.049,
    "args": {
      "External id": 538643,"Record function id": 0, "Ev Idx": 4114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746191.677, "dur": 5.215,
    "args": {
      "External id": 538644,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746200.093, "dur": 0.887,
    "args": {
      "External id": 538645,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746202.957, "dur": 1.110,
    "args": {
      "External id": 538646,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746205.850, "dur": 0.824,
    "args": {
      "External id": 538647,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746208.504, "dur": 1.567,
    "args": {
      "External id": 538648,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746211.604, "dur": 0.954,
    "args": {
      "External id": 538649,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746214.032, "dur": 0.730,
    "args": {
      "External id": 538650,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746216.075, "dur": 0.628,
    "args": {
      "External id": 538651,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746218.083, "dur": 0.812,
    "args": {
      "External id": 538652,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622746220.074, "dur": 0.783,
    "args": {
      "External id": 538653,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622746238.910, "dur": 140.222,
    "args": {
      "External id": 538654,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622746255.267, "dur": 119.666,
    "args": {
      "External id": 538655,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622746267.152, "dur": 14.307,
    "args": {
      "External id": 538656,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622746284.139, "dur": 63.836,
    "args": {
      "External id": 538657,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 4128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622746286.933, "dur": 60.726,
    "args": {
      "External id": 538658,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 4129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746290.867, "dur": 5.484,
    "args": {
      "External id": 538659,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622746298.017, "dur": 49.089,
    "args": {
      "External id": 538660,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 4131
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.1", "pid": 4183438, "tid": 31367,
    "ts": 679622746494.971, "dur": 573.254,
    "args": {
      "External id": 538661,"Record function id": 0, "Ev Idx": 4132
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.1)", "pid": 4183438, "tid": 31367,
    "ts": 679622746510.715, "dur": 545.294,
    "args": {
      "External id": 538662,"Record function id": 0, "Ev Idx": 4133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622746570.091, "dur": 5.728,
    "args": {
      "External id": 538663,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622746594.971, "dur": 24.318,
    "args": {
      "External id": 538664,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746599.077, "dur": 1.700,
    "args": {
      "External id": 538665,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746602.116, "dur": 0.570,
    "args": {
      "External id": 538666,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746603.522, "dur": 0.342,
    "args": {
      "External id": 538667,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746604.768, "dur": 0.604,
    "args": {
      "External id": 538668,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746607.127, "dur": 0.449,
    "args": {
      "External id": 538669,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746608.922, "dur": 0.282,
    "args": {
      "External id": 538670,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746612.042, "dur": 0.596,
    "args": {
      "External id": 538671,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746613.390, "dur": 1.212,
    "args": {
      "External id": 538672,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746615.634, "dur": 0.352,
    "args": {
      "External id": 538673,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622746627.947, "dur": 75.951,
    "args": {
      "External id": 538674,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622746742.565, "dur": 95.425,
    "args": {
      "External id": 538675,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 4146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622746753.312, "dur": 4.558,
    "args": {
      "External id": 538676,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622746763.272, "dur": 9.792,
    "args": {
      "External id": 538677,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622746767.088, "dur": 5.594,
    "args": {
      "External id": 538678,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 4149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746770.545, "dur": 0.632,
    "args": {
      "External id": 538679,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622746781.395, "dur": 17.630,
    "args": {
      "External id": 538680,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746783.462, "dur": 0.420,
    "args": {
      "External id": 538681,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746784.917, "dur": 0.507,
    "args": {
      "External id": 538682,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746786.663, "dur": 0.499,
    "args": {
      "External id": 538683,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746788.493, "dur": 0.592,
    "args": {
      "External id": 538684,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746790.152, "dur": 1.138,
    "args": {
      "External id": 538685,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746792.297, "dur": 0.402,
    "args": {
      "External id": 538686,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746793.670, "dur": 0.386,
    "args": {
      "External id": 538687,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746795.065, "dur": 0.245,
    "args": {
      "External id": 538688,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622746796.528, "dur": 0.229,
    "args": {
      "External id": 538689,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622746809.044, "dur": 21.260,
    "args": {
      "External id": 538690,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622746883.213, "dur": 110.067,
    "args": {
      "External id": 538691,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 4162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622746905.863, "dur": 83.882,
    "args": {
      "External id": 538692,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4163, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622746914.805, "dur": 70.514,
    "args": {
      "External id": 538693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 4164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622747005.458, "dur": 1.803,
    "args": {
      "External id": 538694,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4165, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622747082.732, "dur": 1657.423,
    "args": {
      "External id": 538695,"Sequence number": 6319422, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4166
    }
  },
  {
    "ph": "f", "id": 206, "pid": 4183438, "tid": 31367, "ts": 679622747082.732,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622747213.078, "dur": 102.441,
    "args": {
      "External id": 538696,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622747353.943, "dur": 36.294,
    "args": {
      "External id": 538697,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 4168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622747406.036, "dur": 40.676,
    "args": {
      "External id": 538698,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 4169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622747455.626, "dur": 25.315,
    "args": {
      "External id": 538699,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622747486.889, "dur": 31.472,
    "args": {
      "External id": 538700,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622747525.567, "dur": 19.474,
    "args": {
      "External id": 538701,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622747550.883, "dur": 29.549,
    "args": {
      "External id": 538702,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622747600.691, "dur": 23.110,
    "args": {
      "External id": 538703,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 4174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622747647.404, "dur": 69.645,
    "args": {
      "External id": 538704,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622747739.845, "dur": 18.829,
    "args": {
      "External id": 538705,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622747770.107, "dur": 14.735,
    "args": {
      "External id": 538706,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622747794.106, "dur": 40.867,
    "args": {
      "External id": 538707,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622747838.401, "dur": 30.804,
    "args": {
      "External id": 538708,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622747894.636, "dur": 162.801,
    "args": {
      "External id": 538709,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622747970.155, "dur": 5.945,
    "args": {
      "External id": 538710,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622747977.913, "dur": 2.473,
    "args": {
      "External id": 538711,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622748085.981, "dur": 39.914,
    "args": {
      "External id": 538712,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622748139.530, "dur": 14.957,
    "args": {
      "External id": 538713,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622748163.441, "dur": 47.955,
    "args": {
      "External id": 538714,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622748218.461, "dur": 31.609,
    "args": {
      "External id": 538715,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622748256.799, "dur": 27.669,
    "args": {
      "External id": 538716,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622748288.462, "dur": 28.412,
    "args": {
      "External id": 538717,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622748322.314, "dur": 26.770,
    "args": {
      "External id": 538718,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622748358.848, "dur": 38.117,
    "args": {
      "External id": 538719,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622748421.318, "dur": 31.492,
    "args": {
      "External id": 538720,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 4191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622748476.390, "dur": 32.592,
    "args": {
      "External id": 538721,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622748534.125, "dur": 37.868,
    "args": {
      "External id": 538722,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622748600.002, "dur": 21.004,
    "args": {
      "External id": 538723,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622748637.095, "dur": 58.040,
    "args": {
      "External id": 538724,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 4195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622748959.454, "dur": 18.876,
    "args": {
      "External id": 538725,"Record function id": 0, "Ev Idx": 4196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622748963.255, "dur": 13.806,
    "args": {
      "External id": 538726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622748968.682, "dur": 7.296,
    "args": {
      "External id": 538727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622748970.551, "dur": 5.014,
    "args": {
      "External id": 538728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749165.635, "dur": 10.452,
    "args": {
      "External id": 538729,"Record function id": 0, "Ev Idx": 4200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749168.713, "dur": 6.367,
    "args": {
      "External id": 538730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749170.375, "dur": 2.805,
    "args": {
      "External id": 538731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749171.080, "dur": 1.748,
    "args": {
      "External id": 538732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749182.169, "dur": 5.948,
    "args": {
      "External id": 538733,"Record function id": 0, "Ev Idx": 4204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749183.869, "dur": 3.547,
    "args": {
      "External id": 538734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749184.926, "dur": 1.933,
    "args": {
      "External id": 538735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749185.653, "dur": 1.034,
    "args": {
      "External id": 538736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749192.973, "dur": 7.919,
    "args": {
      "External id": 538737,"Record function id": 0, "Ev Idx": 4208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749194.973, "dur": 5.230,
    "args": {
      "External id": 538738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749196.126, "dur": 3.481,
    "args": {
      "External id": 538739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749196.818, "dur": 2.620,
    "args": {
      "External id": 538740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749205.612, "dur": 6.040,
    "args": {
      "External id": 538741,"Record function id": 0, "Ev Idx": 4212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749207.651, "dur": 3.241,
    "args": {
      "External id": 538742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749208.498, "dur": 1.766,
    "args": {
      "External id": 538743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749209.108, "dur": 0.958,
    "args": {
      "External id": 538744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749216.260, "dur": 5.967,
    "args": {
      "External id": 538745,"Record function id": 0, "Ev Idx": 4216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749218.011, "dur": 3.481,
    "args": {
      "External id": 538746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749218.916, "dur": 1.980,
    "args": {
      "External id": 538747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749219.795, "dur": 0.970,
    "args": {
      "External id": 538748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749227.086, "dur": 6.290,
    "args": {
      "External id": 538749,"Record function id": 0, "Ev Idx": 4220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749229.177, "dur": 3.475,
    "args": {
      "External id": 538750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749230.165, "dur": 1.882,
    "args": {
      "External id": 538751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749231.170, "dur": 0.715,
    "args": {
      "External id": 538752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749238.189, "dur": 6.165,
    "args": {
      "External id": 538753,"Record function id": 0, "Ev Idx": 4224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749240.037, "dur": 3.594,
    "args": {
      "External id": 538754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749240.899, "dur": 2.132,
    "args": {
      "External id": 538755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749241.826, "dur": 1.043,
    "args": {
      "External id": 538756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749249.127, "dur": 4.320,
    "args": {
      "External id": 538757,"Record function id": 0, "Ev Idx": 4228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622749250.422, "dur": 2.603,
    "args": {
      "External id": 538758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749250.988, "dur": 1.494,
    "args": {
      "External id": 538759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622749251.727, "dur": 0.671,
    "args": {
      "External id": 538760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622749258.131, "dur": 15238.656,
    "args": {
      "External id": 538761,"Record function id": 0, "Sequence number": 6319421, "Fwd thread id": 1, "Ev Idx": 4232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622749259.296, "dur": 15229.156,
    "args": {
      "External id": 538762,"Sequence number": 6319421, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4233
    }
  },
  {
    "ph": "f", "id": 207, "pid": 4183438, "tid": 31367, "ts": 679622749259.296,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.2)", "pid": 4183438, "tid": 31367,
    "ts": 679622749292.708, "dur": 41.448,
    "args": {
      "External id": 538763,"Record function id": 0, "Ev Idx": 4234
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.2)", "pid": 4183438, "tid": 31367,
    "ts": 679622749342.320, "dur": 67.713,
    "args": {
      "External id": 538764,"Record function id": 0, "Ev Idx": 4235
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.2)", "pid": 4183438, "tid": 31367,
    "ts": 679622749416.897, "dur": 15063.563,
    "args": {
      "External id": 538765,"Record function id": 0, "Ev Idx": 4236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622749513.116, "dur": 8.516,
    "args": {
      "External id": 538766,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622749531.433, "dur": 5.093,
    "args": {
      "External id": 538767,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622749550.490, "dur": 14221.359,
    "args": {
      "External id": 538768,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622749564.694, "dur": 14196.145,
    "args": {
      "External id": 538769,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622749591.355, "dur": 15.731,
    "args": {
      "External id": 538770,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622749611.043, "dur": 14104.740,
    "args": {
      "External id": 538771,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 4242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622749613.380, "dur": 14101.454,
    "args": {
      "External id": 538772,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 4243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622749617.382, "dur": 5.246,
    "args": {
      "External id": 538773,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622749624.387, "dur": 14085.623,
    "args": {
      "External id": 538774,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 4245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622763879.622, "dur": 12.332,
    "args": {
      "External id": 538775,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 4246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622763883.197, "dur": 8.416,
    "args": {
      "External id": 538776,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622763921.538, "dur": 272.199,
    "args": {
      "External id": 538777,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 4248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622763947.672, "dur": 240.842,
    "args": {
      "External id": 538778,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4249, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622763959.013, "dur": 223.747,
    "args": {
      "External id": 538779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 4250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622764211.085, "dur": 2.214,
    "args": {
      "External id": 538780,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4251, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764274.830, "dur": 6.856,
    "args": {
      "External id": 538781,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764328.323, "dur": 1.540,
    "args": {
      "External id": 538782,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764346.291, "dur": 1.512,
    "args": {
      "External id": 538783,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764360.253, "dur": 1.990,
    "args": {
      "External id": 538784,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764374.258, "dur": 0.947,
    "args": {
      "External id": 538785,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764385.488, "dur": 0.815,
    "args": {
      "External id": 538786,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764395.994, "dur": 1.072,
    "args": {
      "External id": 538787,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764406.932, "dur": 1.776,
    "args": {
      "External id": 538788,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764417.169, "dur": 0.628,
    "args": {
      "External id": 538789,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622764510.189, "dur": 3109.383,
    "args": {
      "External id": 538790,"Record function id": 0, "Ev Idx": 4261
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.1)", "pid": 4183438, "tid": 31367,
    "ts": 679622764529.585, "dur": 1022.573,
    "args": {
      "External id": 538791,"Record function id": 0, "Ev Idx": 4262
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.1)", "pid": 4183438, "tid": 31367,
    "ts": 679622764547.198, "dur": 348.639,
    "args": {
      "External id": 538792,"Record function id": 0, "Ev Idx": 4263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764629.693, "dur": 4.355,
    "args": {
      "External id": 538793,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764637.354, "dur": 0.881,
    "args": {
      "External id": 538794,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764640.034, "dur": 0.772,
    "args": {
      "External id": 538795,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764642.673, "dur": 0.825,
    "args": {
      "External id": 538796,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764644.924, "dur": 1.050,
    "args": {
      "External id": 538797,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764647.444, "dur": 0.897,
    "args": {
      "External id": 538798,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764688.487, "dur": 2.094,
    "args": {
      "External id": 538799,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764694.971, "dur": 1.010,
    "args": {
      "External id": 538800,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764697.586, "dur": 0.799,
    "args": {
      "External id": 538801,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622764699.722, "dur": 0.813,
    "args": {
      "External id": 538802,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622764719.344, "dur": 146.835,
    "args": {
      "External id": 538803,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622764735.590, "dur": 126.160,
    "args": {
      "External id": 538804,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622764750.609, "dur": 14.222,
    "args": {
      "External id": 538805,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622764767.552, "dur": 65.690,
    "args": {
      "External id": 538806,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 4277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622764770.038, "dur": 62.922,
    "args": {
      "External id": 538807,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 4278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622764773.864, "dur": 6.328,
    "args": {
      "External id": 538808,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622764783.649, "dur": 48.444,
    "args": {
      "External id": 538809,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 4280
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.0", "pid": 4183438, "tid": 31367,
    "ts": 679622764979.819, "dur": 565.193,
    "args": {
      "External id": 538810,"Record function id": 0, "Ev Idx": 4281
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.0)", "pid": 4183438, "tid": 31367,
    "ts": 679622764995.684, "dur": 538.094,
    "args": {
      "External id": 538811,"Record function id": 0, "Ev Idx": 4282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622765057.558, "dur": 5.719,
    "args": {
      "External id": 538812,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622765077.589, "dur": 41.472,
    "args": {
      "External id": 538813,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765081.743, "dur": 1.687,
    "args": {
      "External id": 538814,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765085.043, "dur": 0.568,
    "args": {
      "External id": 538815,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765086.841, "dur": 0.517,
    "args": {
      "External id": 538816,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765088.316, "dur": 0.708,
    "args": {
      "External id": 538817,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765090.497, "dur": 0.705,
    "args": {
      "External id": 538818,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765092.732, "dur": 0.678,
    "args": {
      "External id": 538819,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765094.790, "dur": 0.567,
    "args": {
      "External id": 538820,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765096.700, "dur": 0.903,
    "args": {
      "External id": 538821,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765098.801, "dur": 16.543,
    "args": {
      "External id": 538822,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622765131.341, "dur": 37.043,
    "args": {
      "External id": 538823,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 31367,
    "ts": 679622765206.245, "dur": 100.437,
    "args": {
      "External id": 538824,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 4295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622765216.129, "dur": 4.832,
    "args": {
      "External id": 538825,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 31367,
    "ts": 679622765226.301, "dur": 10.517,
    "args": {
      "External id": 538826,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 31367,
    "ts": 679622765230.561, "dur": 5.856,
    "args": {
      "External id": 538827,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 4298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765234.230, "dur": 1.122,
    "args": {
      "External id": 538828,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 31367,
    "ts": 679622765243.593, "dur": 26.502,
    "args": {
      "External id": 538829,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765245.507, "dur": 0.637,
    "args": {
      "External id": 538830,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765247.232, "dur": 0.524,
    "args": {
      "External id": 538831,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765249.180, "dur": 0.707,
    "args": {
      "External id": 538832,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765250.942, "dur": 4.567,
    "args": {
      "External id": 538833,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765256.700, "dur": 0.869,
    "args": {
      "External id": 538834,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765258.750, "dur": 1.200,
    "args": {
      "External id": 538835,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765260.913, "dur": 0.705,
    "args": {
      "External id": 538836,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765262.800, "dur": 0.604,
    "args": {
      "External id": 538837,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622765264.728, "dur": 0.545,
    "args": {
      "External id": 538838,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622765279.803, "dur": 19.753,
    "args": {
      "External id": 538839,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622765353.420, "dur": 117.410,
    "args": {
      "External id": 538840,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 4311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622765380.857, "dur": 86.507,
    "args": {
      "External id": 538841,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4312, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 31367,
    "ts": 679622765390.966, "dur": 72.277,
    "args": {
      "External id": 538842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 4313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622765484.946, "dur": 1.927,
    "args": {
      "External id": 538843,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4314, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622765558.900, "dur": 2039.868,
    "args": {
      "External id": 538844,"Sequence number": 6319420, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4315
    }
  },
  {
    "ph": "f", "id": 208, "pid": 4183438, "tid": 31367, "ts": 679622765558.900,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622765710.865, "dur": 104.622,
    "args": {
      "External id": 538845,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622765855.668, "dur": 38.477,
    "args": {
      "External id": 538846,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 4317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622765911.029, "dur": 40.941,
    "args": {
      "External id": 538847,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 4318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622765961.811, "dur": 25.185,
    "args": {
      "External id": 538848,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622765993.268, "dur": 32.428,
    "args": {
      "External id": 538849,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766032.268, "dur": 20.087,
    "args": {
      "External id": 538850,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766060.413, "dur": 27.919,
    "args": {
      "External id": 538851,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622766127.809, "dur": 24.758,
    "args": {
      "External id": 538852,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 4323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622766171.089, "dur": 26.655,
    "args": {
      "External id": 538853,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622766217.333, "dur": 18.620,
    "args": {
      "External id": 538854,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622766249.224, "dur": 14.133,
    "args": {
      "External id": 538855,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766271.584, "dur": 39.250,
    "args": {
      "External id": 538856,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766314.514, "dur": 30.096,
    "args": {
      "External id": 538857,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622766370.320, "dur": 168.603,
    "args": {
      "External id": 538858,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622766447.392, "dur": 6.209,
    "args": {
      "External id": 538859,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622766455.492, "dur": 3.355,
    "args": {
      "External id": 538860,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622766567.972, "dur": 22.496,
    "args": {
      "External id": 538861,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622766602.061, "dur": 14.253,
    "args": {
      "External id": 538862,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766623.770, "dur": 78.088,
    "args": {
      "External id": 538863,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766710.293, "dur": 34.977,
    "args": {
      "External id": 538864,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766753.446, "dur": 28.778,
    "args": {
      "External id": 538865,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766789.413, "dur": 32.277,
    "args": {
      "External id": 538866,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766831.358, "dur": 37.264,
    "args": {
      "External id": 538867,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622766879.826, "dur": 41.697,
    "args": {
      "External id": 538868,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622766948.295, "dur": 48.809,
    "args": {
      "External id": 538869,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 4340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622767035.312, "dur": 36.460,
    "args": {
      "External id": 538870,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622767092.208, "dur": 42.262,
    "args": {
      "External id": 538871,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622767330.218, "dur": 187.554,
    "args": {
      "External id": 538872,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622767535.988, "dur": 24.476,
    "args": {
      "External id": 538873,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 4344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767642.954, "dur": 49.422,
    "args": {
      "External id": 538874,"Record function id": 0, "Ev Idx": 4345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767646.266, "dur": 44.430,
    "args": {
      "External id": 538875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767682.104, "dur": 7.233,
    "args": {
      "External id": 538876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767684.036, "dur": 5.013,
    "args": {
      "External id": 538877,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767698.353, "dur": 6.119,
    "args": {
      "External id": 538878,"Record function id": 0, "Ev Idx": 4349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767700.690, "dur": 3.404,
    "args": {
      "External id": 538879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767701.939, "dur": 1.695,
    "args": {
      "External id": 538880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767702.501, "dur": 1.033,
    "args": {
      "External id": 538881,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767707.670, "dur": 4.229,
    "args": {
      "External id": 538882,"Record function id": 0, "Ev Idx": 4353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767708.890, "dur": 2.653,
    "args": {
      "External id": 538883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767709.400, "dur": 1.733,
    "args": {
      "External id": 538884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767710.028, "dur": 0.972,
    "args": {
      "External id": 538885,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767714.969, "dur": 4.353,
    "args": {
      "External id": 538886,"Record function id": 0, "Ev Idx": 4357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767716.360, "dur": 2.587,
    "args": {
      "External id": 538887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767716.986, "dur": 1.497,
    "args": {
      "External id": 538888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767717.427, "dur": 0.961,
    "args": {
      "External id": 538889,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767722.261, "dur": 3.956,
    "args": {
      "External id": 538890,"Record function id": 0, "Ev Idx": 4361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767723.399, "dur": 2.454,
    "args": {
      "External id": 538891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767723.873, "dur": 1.527,
    "args": {
      "External id": 538892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767724.210, "dur": 1.097,
    "args": {
      "External id": 538893,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767729.181, "dur": 4.140,
    "args": {
      "External id": 538894,"Record function id": 0, "Ev Idx": 4365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767730.542, "dur": 2.409,
    "args": {
      "External id": 538895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767731.000, "dur": 1.497,
    "args": {
      "External id": 538896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767731.594, "dur": 0.809,
    "args": {
      "External id": 538897,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767736.400, "dur": 4.213,
    "args": {
      "External id": 538898,"Record function id": 0, "Ev Idx": 4369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767737.748, "dur": 2.492,
    "args": {
      "External id": 538899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767738.172, "dur": 1.610,
    "args": {
      "External id": 538900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767738.879, "dur": 0.834,
    "args": {
      "External id": 538901,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767743.693, "dur": 3.668,
    "args": {
      "External id": 538902,"Record function id": 0, "Ev Idx": 4373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767744.911, "dur": 2.089,
    "args": {
      "External id": 538903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767745.328, "dur": 1.251,
    "args": {
      "External id": 538904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767745.805, "dur": 0.680,
    "args": {
      "External id": 538905,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767750.674, "dur": 4.068,
    "args": {
      "External id": 538906,"Record function id": 0, "Ev Idx": 4377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622767751.819, "dur": 2.550,
    "args": {
      "External id": 538907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767752.277, "dur": 1.659,
    "args": {
      "External id": 538908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622767752.940, "dur": 0.901,
    "args": {
      "External id": 538909,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622767758.584, "dur": 15083.813,
    "args": {
      "External id": 538910,"Record function id": 0, "Sequence number": 6319419, "Fwd thread id": 1, "Ev Idx": 4381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622767760.034, "dur": 15073.755,
    "args": {
      "External id": 538911,"Sequence number": 6319419, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4382
    }
  },
  {
    "ph": "f", "id": 209, "pid": 4183438, "tid": 31367, "ts": 679622767760.034,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.1)", "pid": 4183438, "tid": 31367,
    "ts": 679622767791.937, "dur": 41.212,
    "args": {
      "External id": 538912,"Record function id": 0, "Ev Idx": 4383
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.1)", "pid": 4183438, "tid": 31367,
    "ts": 679622767841.074, "dur": 67.723,
    "args": {
      "External id": 538913,"Record function id": 0, "Ev Idx": 4384
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.1)", "pid": 4183438, "tid": 31367,
    "ts": 679622767915.384, "dur": 14910.469,
    "args": {
      "External id": 538914,"Record function id": 0, "Ev Idx": 4385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622768009.526, "dur": 8.101,
    "args": {
      "External id": 538915,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622768027.515, "dur": 5.012,
    "args": {
      "External id": 538916,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622768046.552, "dur": 14028.982,
    "args": {
      "External id": 538917,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622768060.622, "dur": 14003.548,
    "args": {
      "External id": 538918,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622768091.838, "dur": 30.349,
    "args": {
      "External id": 538919,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622768127.103, "dur": 13897.081,
    "args": {
      "External id": 538920,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 4391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622768130.382, "dur": 13892.831,
    "args": {
      "External id": 538921,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 4392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622768134.280, "dur": 6.269,
    "args": {
      "External id": 538922,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622768142.313, "dur": 13876.551,
    "args": {
      "External id": 538923,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 4394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622782191.864, "dur": 12.247,
    "args": {
      "External id": 538924,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 4395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622782195.302, "dur": 8.248,
    "args": {
      "External id": 538925,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622782236.592, "dur": 275.458,
    "args": {
      "External id": 538926,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 4397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622782263.771, "dur": 240.871,
    "args": {
      "External id": 538927,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4398, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622782279.904, "dur": 218.397,
    "args": {
      "External id": 538928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 4399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622782527.879, "dur": 2.178,
    "args": {
      "External id": 538929,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4400, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782583.017, "dur": 6.670,
    "args": {
      "External id": 538930,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782633.744, "dur": 1.857,
    "args": {
      "External id": 538931,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782683.642, "dur": 2.599,
    "args": {
      "External id": 538932,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782704.798, "dur": 1.027,
    "args": {
      "External id": 538933,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782717.151, "dur": 1.229,
    "args": {
      "External id": 538934,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782728.840, "dur": 1.294,
    "args": {
      "External id": 538935,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782741.736, "dur": 1.209,
    "args": {
      "External id": 538936,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782753.419, "dur": 1.128,
    "args": {
      "External id": 538937,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622782764.419, "dur": 0.878,
    "args": {
      "External id": 538938,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622782857.800, "dur": 2038.982,
    "args": {
      "External id": 538939,"Record function id": 0, "Ev Idx": 4410
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.0)", "pid": 4183438, "tid": 31367,
    "ts": 679622782876.456, "dur": 439.218,
    "args": {
      "External id": 538940,"Record function id": 0, "Ev Idx": 4411
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.0)", "pid": 4183438, "tid": 31367,
    "ts": 679622782890.555, "dur": 326.770,
    "args": {
      "External id": 538941,"Record function id": 0, "Ev Idx": 4412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622782974.481, "dur": 4.423,
    "args": {
      "External id": 538942,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622782985.188, "dur": 1.113,
    "args": {
      "External id": 538943,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622782988.106, "dur": 0.903,
    "args": {
      "External id": 538944,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622782990.788, "dur": 1.041,
    "args": {
      "External id": 538945,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622782993.594, "dur": 1.066,
    "args": {
      "External id": 538946,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622782996.013, "dur": 0.963,
    "args": {
      "External id": 538947,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622782998.530, "dur": 1.038,
    "args": {
      "External id": 538948,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622783001.298, "dur": 0.923,
    "args": {
      "External id": 538949,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622783004.188, "dur": 1.250,
    "args": {
      "External id": 538950,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622783006.966, "dur": 1.461,
    "args": {
      "External id": 538951,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622783025.345, "dur": 161.896,
    "args": {
      "External id": 538952,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622783045.139, "dur": 137.370,
    "args": {
      "External id": 538953,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622783057.879, "dur": 13.901,
    "args": {
      "External id": 538954,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622783074.745, "dur": 78.927,
    "args": {
      "External id": 538955,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 4426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622783078.745, "dur": 74.558,
    "args": {
      "External id": 538956,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 4427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622783082.974, "dur": 5.420,
    "args": {
      "External id": 538957,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622783090.107, "dur": 62.418,
    "args": {
      "External id": 538958,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 4429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622783323.485, "dur": 1551.373,
    "args": {
      "External id": 538959,"Sequence number": 6319418, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4430
    }
  },
  {
    "ph": "f", "id": 210, "pid": 4183438, "tid": 31367, "ts": 679622783323.485,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622783430.759, "dur": 105.912,
    "args": {
      "External id": 538960,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [2816, 1], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 4183438, "tid": 31367,
    "ts": 679622783571.982, "dur": 35.350,
    "args": {
      "External id": 538961,"kernel_hash": "c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/4p/c4po4gnedw4lkxajnn6p3wyqqi24sdacil3q3sdl47drs4et3op3.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [65536, 2816], [16, 4096, 2816], [16, 4096, 2816], [16, 4096, 2816], []], "Ev Idx": 4432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 4183438, "tid": 31367,
    "ts": 679622783622.891, "dur": 81.073,
    "args": {
      "External id": 538962,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 1024], [0, 2816, 1], [2883584, 2816, 1]], "Input Dims": [[1, 1024, 65536], [1, 65536, 2816], [1, 1024, 2816]], "Ev Idx": 4433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622783718.067, "dur": 29.459,
    "args": {
      "External id": 538963,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622783754.059, "dur": 33.109,
    "args": {
      "External id": 538964,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622783794.993, "dur": 20.946,
    "args": {
      "External id": 538965,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2816], [1024, 1], [1024, 1]], "Input Dims": [[2816, 65536], [65536, 1024], [2816, 1024]], "Ev Idx": 4436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622783824.250, "dur": 29.615,
    "args": {
      "External id": 538966,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 4183438, "tid": 31367,
    "ts": 679622783879.167, "dur": 27.120,
    "args": {
      "External id": 538967,"kernel_hash": "cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/oh/cohoku7zuxcsdfkchrrz4kspw7usafsc5d3yc5qjdd22veuugeck.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 4438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 4183438, "tid": 31367,
    "ts": 679622783923.892, "dur": 32.943,
    "args": {
      "External id": 538968,"kernel_hash": "clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/lm/clm7bfnx7pnl5oqs77dih4ccwole4f5pdm4hil7p5ob7aogwgcwv.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622783976.541, "dur": 18.437,
    "args": {
      "External id": 538969,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622784007.579, "dur": 15.283,
    "args": {
      "External id": 538970,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784030.994, "dur": 38.233,
    "args": {
      "External id": 538971,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784072.622, "dur": 46.570,
    "args": {
      "External id": 538972,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 4183438, "tid": 31367,
    "ts": 679622784149.686, "dur": 178.163,
    "args": {
      "External id": 538973,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [65536, 4096, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 16, 4096], [16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622784234.588, "dur": 6.777,
    "args": {
      "External id": 538974,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622784243.130, "dur": 2.888,
    "args": {
      "External id": 538975,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622784358.187, "dur": 23.783,
    "args": {
      "External id": 538976,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 31367,
    "ts": 679622784393.259, "dur": 15.782,
    "args": {
      "External id": 538977,"kernel_hash": "cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/xy/cxy3rbhc5coe5hycpj4oiqwzfsszxnphpxg4hr2bzokiu5chquvc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784417.731, "dur": 48.237,
    "args": {
      "External id": 538978,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784471.862, "dur": 31.923,
    "args": {
      "External id": 538979,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784510.252, "dur": 27.705,
    "args": {
      "External id": 538980,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784543.246, "dur": 28.313,
    "args": {
      "External id": 538981,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784576.755, "dur": 26.326,
    "args": {
      "External id": 538982,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 1024], [1024, 1], [1024, 1]], "Input Dims": [[1024, 65536], [65536, 1024], [1024, 1024]], "Ev Idx": 4453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 31367,
    "ts": 679622784610.483, "dur": 28.934,
    "args": {
      "External id": 538983,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 4183438, "tid": 31367,
    "ts": 679622784688.752, "dur": 25.560,
    "args": {
      "External id": 538984,"kernel_hash": "cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxquaruus47mvrquq3gfsj3is6jdepxek5jt3bsov27xnlnlypmz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [65536, 1024], []], "Ev Idx": 4455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 4183438, "tid": 31367,
    "ts": 679622784732.904, "dur": 23.795,
    "args": {
      "External id": 538985,"kernel_hash": "chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "1024", "1", "497", "132", "True", "1024", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/hl/chlkjt3hicxy6k2bpahxa6z5u7fh2potkgg72keqyi6pjjgwpucm.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [1024], [65536, 1024], [65536, 1024], [132, 1024], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 4183438, "tid": 31367,
    "ts": 679622784770.646, "dur": 16.954,
    "args": {
      "External id": 538986,"kernel_hash": "cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "66"], "kernel_file": "/tmp/torchinductor_cvm/yz/cyztcfbseh2txumvnyll7ahf6t5d5tqi6dmacahyih75kq7cgxcr.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [2048, 1, 1024], [], []], "Input Dims": [[132, 1024], [1, 1024, 2], [], []], "Ev Idx": 4457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 4183438, "tid": 31367,
    "ts": 679622784802.407, "dur": 15.178,
    "args": {
      "External id": 538987,"kernel_hash": "cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y", "grid": "grid(1024,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "1024", "2"], "kernel_file": "/tmp/torchinductor_cvm/th/cthfckwcw2pzhsppv2sbhauqwtg23csgbsalqrbuub6hj6pj4p7y.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1, 1024], [1024, 1], [], []], "Input Dims": [[1, 1024, 2], [1, 1024], [], []], "Ev Idx": 4458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 4183438, "tid": 31367,
    "ts": 679622784828.889, "dur": 17.130,
    "args": {
      "External id": 538988,"kernel_hash": "csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/sq/csqhbcpoi4yxxvcel4c4wfiuwwaxltg3nqs73dgwqy7o36qde5hr.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], [16, 4096, 1024], []], "Ev Idx": 4459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784919.346, "dur": 15.859,
    "args": {
      "External id": 538989,"Record function id": 0, "Ev Idx": 4460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784922.502, "dur": 11.777,
    "args": {
      "External id": 538990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784926.926, "dur": 6.450,
    "args": {
      "External id": 538991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784928.406, "dur": 4.834,
    "args": {
      "External id": 538992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784939.004, "dur": 4.970,
    "args": {
      "External id": 538993,"Record function id": 0, "Ev Idx": 4464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784940.444, "dur": 3.031,
    "args": {
      "External id": 538994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784941.489, "dur": 1.538,
    "args": {
      "External id": 538995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784942.053, "dur": 0.884,
    "args": {
      "External id": 538996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784947.217, "dur": 5.115,
    "args": {
      "External id": 538997,"Record function id": 0, "Ev Idx": 4468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784949.106, "dur": 2.833,
    "args": {
      "External id": 538998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784949.764, "dur": 1.791,
    "args": {
      "External id": 538999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784950.352, "dur": 1.112,
    "args": {
      "External id": 539000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1024]], "Ev Idx": 4471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784955.495, "dur": 4.843,
    "args": {
      "External id": 539001,"Record function id": 0, "Ev Idx": 4472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784957.191, "dur": 2.757,
    "args": {
      "External id": 539002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784958.226, "dur": 1.311,
    "args": {
      "External id": 539003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784958.593, "dur": 0.851,
    "args": {
      "External id": 539004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 4475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784963.424, "dur": 4.544,
    "args": {
      "External id": 539005,"Record function id": 0, "Ev Idx": 4476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784965.059, "dur": 2.497,
    "args": {
      "External id": 539006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784965.618, "dur": 1.556,
    "args": {
      "External id": 539007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784966.005, "dur": 1.075,
    "args": {
      "External id": 539008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784971.078, "dur": 4.344,
    "args": {
      "External id": 539009,"Record function id": 0, "Ev Idx": 4480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784972.439, "dur": 2.604,
    "args": {
      "External id": 539010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784973.245, "dur": 1.410,
    "args": {
      "External id": 539011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784973.755, "dur": 0.810,
    "args": {
      "External id": 539012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784978.597, "dur": 3.938,
    "args": {
      "External id": 539013,"Record function id": 0, "Ev Idx": 4484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784979.905, "dur": 2.193,
    "args": {
      "External id": 539014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784980.364, "dur": 1.338,
    "args": {
      "External id": 539015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784980.951, "dur": 0.673,
    "args": {
      "External id": 539016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784985.784, "dur": 4.111,
    "args": {
      "External id": 539017,"Record function id": 0, "Ev Idx": 4488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784987.318, "dur": 2.191,
    "args": {
      "External id": 539018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784987.844, "dur": 1.280,
    "args": {
      "External id": 539019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784988.363, "dur": 0.667,
    "args": {
      "External id": 539020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 4491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784993.168, "dur": 3.512,
    "args": {
      "External id": 539021,"Record function id": 0, "Ev Idx": 4492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622784994.232, "dur": 2.038,
    "args": {
      "External id": 539022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784994.695, "dur": 1.184,
    "args": {
      "External id": 539023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622784995.058, "dur": 0.728,
    "args": {
      "External id": 539024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 4495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622785000.938, "dur": 16161.128,
    "args": {
      "External id": 539025,"Record function id": 0, "Sequence number": 6319417, "Fwd thread id": 1, "Ev Idx": 4496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622785002.086, "dur": 16150.671,
    "args": {
      "External id": 539026,"Sequence number": 6319417, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4497
    }
  },
  {
    "ph": "f", "id": 211, "pid": 4183438, "tid": 31367, "ts": 679622785002.086,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.0)", "pid": 4183438, "tid": 31367,
    "ts": 679622785031.187, "dur": 41.613,
    "args": {
      "External id": 539027,"Record function id": 0, "Ev Idx": 4498
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.0)", "pid": 4183438, "tid": 31367,
    "ts": 679622785080.424, "dur": 85.846,
    "args": {
      "External id": 539028,"Record function id": 0, "Ev Idx": 4499
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.0)", "pid": 4183438, "tid": 31367,
    "ts": 679622785174.537, "dur": 15969.467,
    "args": {
      "External id": 539029,"Record function id": 0, "Ev Idx": 4500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622785290.433, "dur": 8.696,
    "args": {
      "External id": 539030,"Record function id": 0, "Concrete Inputs": ["[12847104]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622785312.722, "dur": 6.932,
    "args": {
      "External id": 539031,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622785337.924, "dur": 15006.354,
    "args": {
      "External id": 539032,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622785372.167, "dur": 14960.967,
    "args": {
      "External id": 539033,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [1024, 1], [1024, 1], [1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], [], [], [1605888, 1]], "Input Dims": [[[1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], [], [], [8, 1605888]], "Ev Idx": 4504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622785404.609, "dur": 20.583,
    "args": {
      "External id": 539034,"Record function id": 0, "Concrete Inputs": ["[3194]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622785430.221, "dur": 14858.343,
    "args": {
      "External id": 539035,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], [], []], "Ev Idx": 4506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622785433.538, "dur": 14853.990,
    "args": {
      "External id": 539036,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3194], [], [], [], [], [], []], "Ev Idx": 4507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622785438.733, "dur": 9.821,
    "args": {
      "External id": 539037,"Record function id": 0, "Concrete Inputs": ["[3194]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622785451.089, "dur": 14831.668,
    "args": {
      "External id": 539038,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3194], [3194], []], "Ev Idx": 4509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622800457.222, "dur": 13.224,
    "args": {
      "External id": 539039,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[12847104], [], [], [], [], []], "Ev Idx": 4510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622800460.961, "dur": 9.135,
    "args": {
      "External id": 539040,"Record function id": 0, "Concrete Inputs": ["[1605888]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622800503.772, "dur": 326.566,
    "args": {
      "External id": 539041,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[1605888], [12847104], [], [], [], []], "Ev Idx": 4512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622800531.397, "dur": 293.612,
    "args": {
      "External id": 539042,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1605888, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[12847104], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4513, "In msg nelems": 12847104
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622800544.592, "dur": 274.747,
    "args": {
      "External id": 539043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12847104]], "Ev Idx": 4514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622800850.815, "dur": 2.322,
    "args": {
      "External id": 539044,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4515, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622800914.420, "dur": 7.020,
    "args": {
      "External id": 539045,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622800966.889, "dur": 1.786,
    "args": {
      "External id": 539046,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622800985.131, "dur": 1.392,
    "args": {
      "External id": 539047,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "131200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622800999.205, "dur": 0.952,
    "args": {
      "External id": 539048,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "262272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622801012.100, "dur": 0.898,
    "args": {
      "External id": 539049,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "393344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622801023.173, "dur": 1.090,
    "args": {
      "External id": 539050,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622801036.140, "dur": 0.981,
    "args": {
      "External id": 539051,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622801048.292, "dur": 1.189,
    "args": {
      "External id": 539052,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "884992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622801059.424, "dur": 1.028,
    "args": {
      "External id": 539053,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "1245440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622801178.543, "dur": 285.044,
    "args": {
      "External id": 539054,"Record function id": 0, "Sequence number": 6319416, "Fwd thread id": 1, "Ev Idx": 4525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 4183438, "tid": 31367,
    "ts": 679622801181.130, "dur": 273.942,
    "args": {
      "External id": 539055,"Sequence number": 6319416, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4526
    }
  },
  {
    "ph": "f", "id": 212, "pid": 4183438, "tid": 31367, "ts": 679622801181.130,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 4183438, "tid": 31367,
    "ts": 679622801303.506, "dur": 51.285,
    "args": {
      "External id": 539056,"kernel_hash": "c62a63oyn7avzuzdtrtz4qdvhfnops7fz2f53gxtomcyzc7spley", "grid": "grid(32768000,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "32768000"], "kernel_file": "/tmp/torchinductor_cvm/62/c62a63oyn7avzuzdtrtz4qdvhfnops7fz2f53gxtomcyzc7spley.py", "kernel_backend": "triton", "Input type": ["float", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[32000, 1024], []], "Ev Idx": 4527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 4183438, "tid": 31367,
    "ts": 679622801370.945, "dur": 32.430,
    "args": {
      "External id": 539057,"kernel_hash": "cerfxu7bhbn6ulmm7kpiu2toaha32yw6zzr2afv5vcf2lszoj3v4", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/er/cerfxu7bhbn6ulmm7kpiu2toaha32yw6zzr2afv5vcf2lszoj3v4.py", "kernel_backend": "triton", "Input type": ["long int", "c10::BFloat16", "float", "Scalar"], "Input Strides": [[4096, 1], [4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096], [16, 4096, 1024], [32000, 1024], []], "Ev Idx": 4528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_dense_backward_2", "pid": 4183438, "tid": 31367,
    "ts": 679622801419.036, "dur": 21.045,
    "args": {
      "External id": 539058,"kernel_hash": "cny6en3gok6pbitjqamstlopano7gsqxbj5xiwxezt7bsp54rmil", "grid": "grid(32768000,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "32768000"], "kernel_file": "/tmp/torchinductor_cvm/ny/cny6en3gok6pbitjqamstlopano7gsqxbj5xiwxezt7bsp54rmil.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 4529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622801473.249, "dur": 14.968,
    "args": {
      "External id": 539059,"Record function id": 0, "Ev Idx": 4530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 4183438, "tid": 31367,
    "ts": 679622801476.699, "dur": 10.766,
    "args": {
      "External id": 539060,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 4531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 31367,
    "ts": 679622801480.270, "dur": 6.325,
    "args": {
      "External id": 539061,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 4532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 31367,
    "ts": 679622801481.672, "dur": 4.793,
    "args": {
      "External id": 539062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 4533
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::root_post_backward_callback", "pid": 4183438, "tid": 31367,
    "ts": 679622801510.090, "dur": 6657.925,
    "args": {
      "External id": 539063,"Record function id": 0, "Ev Idx": 4534
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate", "pid": 4183438, "tid": 31367,
    "ts": 679622801527.652, "dur": 35.475,
    "args": {
      "External id": 539064,"Record function id": 0, "Ev Idx": 4535
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard", "pid": 4183438, "tid": 31367,
    "ts": 679622801569.189, "dur": 253.568,
    "args": {
      "External id": 539065,"Record function id": 0, "Ev Idx": 4536
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce", "pid": 4183438, "tid": 31367,
    "ts": 679622801831.563, "dur": 6116.572,
    "args": {
      "External id": 539066,"Record function id": 0, "Ev Idx": 4537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622801961.484, "dur": 11.623,
    "args": {
      "External id": 539067,"Record function id": 0, "Concrete Inputs": ["[116925440]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 31367,
    "ts": 679622801984.316, "dur": 5.547,
    "args": {
      "External id": 539068,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[116925440], []], "Ev Idx": 4539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622802008.736, "dur": 4727.990,
    "args": {
      "External id": 539069,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[], [], [], [14615680, 1]], "Input Dims": [[], [], [], [8, 14615680]], "Ev Idx": 4540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 4183438, "tid": 31367,
    "ts": 679622802030.389, "dur": 4692.641,
    "args": {
      "External id": 539070,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[], [], [], [14615680, 1]], "Input Dims": [[], [], [], [8, 14615680]], "Ev Idx": 4541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622802159.439, "dur": 14.281,
    "args": {
      "External id": 539071,"Record function id": 0, "Concrete Inputs": ["[28789]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 31367,
    "ts": 679622802206.851, "dur": 4442.729,
    "args": {
      "External id": 539072,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[28789], [], [], [], [], [], [], []], "Ev Idx": 4543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 31367,
    "ts": 679622802210.459, "dur": 4438.005,
    "args": {
      "External id": 539073,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[28789], [], [], [], [], [], []], "Ev Idx": 4544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622802214.147, "dur": 7.732,
    "args": {
      "External id": 539074,"Record function id": 0, "Concrete Inputs": ["[28789]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 31367,
    "ts": 679622802223.856, "dur": 4418.788,
    "args": {
      "External id": 539075,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[28789], [28789], []], "Ev Idx": 4546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 31367,
    "ts": 679622806870.844, "dur": 12.994,
    "args": {
      "External id": 539076,"Record function id": 0, "Concrete Inputs": ["", "[14615680]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[116925440], [], [], [], [], []], "Ev Idx": 4547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 31367,
    "ts": 679622806874.521, "dur": 8.992,
    "args": {
      "External id": 539077,"Record function id": 0, "Concrete Inputs": ["[14615680]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 4183438, "tid": 31367,
    "ts": 679622806913.091, "dur": 373.723,
    "args": {
      "External id": 539078,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[14615680], [116925440], [], [], [], []], "Ev Idx": 4549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622806940.378, "dur": 341.070,
    "args": {
      "External id": 539079,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 14615680, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[116925440], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4550, "In msg nelems": 116925440
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 4183438, "tid": 31367,
    "ts": 679622806952.763, "dur": 322.768,
    "args": {
      "External id": 539080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[116925440]], "Ev Idx": 4551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 31367,
    "ts": 679622807309.948, "dur": 2.734,
    "args": {
      "External id": 539081,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4552, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807373.644, "dur": 7.216,
    "args": {
      "External id": 539082,"Record function id": 0, "Concrete Inputs": ["", "[4000, 1024]", "[1024, 1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807423.717, "dur": 2.066,
    "args": {
      "External id": 539083,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "4096000"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807440.729, "dur": 2.166,
    "args": {
      "External id": 539084,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "4096128"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807454.445, "dur": 0.947,
    "args": {
      "External id": 539085,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "4227200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807465.643, "dur": 1.099,
    "args": {
      "External id": 539086,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "4358272"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807477.509, "dur": 1.113,
    "args": {
      "External id": 539087,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "4489344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807487.575, "dur": 1.188,
    "args": {
      "External id": 539088,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "4620416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807497.978, "dur": 1.280,
    "args": {
      "External id": 539089,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "4620544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807509.189, "dur": 1.259,
    "args": {
      "External id": 539090,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "4980992"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807520.463, "dur": 1.151,
    "args": {
      "External id": 539091,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "5341440"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807532.404, "dur": 1.116,
    "args": {
      "External id": 539092,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "5701888"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807542.861, "dur": 1.206,
    "args": {
      "External id": 539093,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "5702016"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807553.742, "dur": 1.349,
    "args": {
      "External id": 539094,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "5833088"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807565.213, "dur": 1.153,
    "args": {
      "External id": 539095,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "5964160"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807575.414, "dur": 1.252,
    "args": {
      "External id": 539096,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "6095232"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807585.839, "dur": 1.673,
    "args": {
      "External id": 539097,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "6226304"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807597.625, "dur": 1.228,
    "args": {
      "External id": 539098,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "6226432"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807609.464, "dur": 1.071,
    "args": {
      "External id": 539099,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "6586880"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807621.477, "dur": 1.205,
    "args": {
      "External id": 539100,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "6947328"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807632.270, "dur": 0.928,
    "args": {
      "External id": 539101,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "7307776"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807643.718, "dur": 1.023,
    "args": {
      "External id": 539102,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "7307904"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807691.910, "dur": 2.817,
    "args": {
      "External id": 539103,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "7438976"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807707.030, "dur": 1.401,
    "args": {
      "External id": 539104,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "7570048"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807718.969, "dur": 1.132,
    "args": {
      "External id": 539105,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "7701120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807729.596, "dur": 1.393,
    "args": {
      "External id": 539106,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "7832192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807740.567, "dur": 1.209,
    "args": {
      "External id": 539107,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "7832320"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807751.831, "dur": 1.236,
    "args": {
      "External id": 539108,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "8192768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807763.469, "dur": 1.998,
    "args": {
      "External id": 539109,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "8553216"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807774.942, "dur": 1.333,
    "args": {
      "External id": 539110,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "8913664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807786.368, "dur": 1.194,
    "args": {
      "External id": 539111,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "8913792"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807796.209, "dur": 1.236,
    "args": {
      "External id": 539112,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "9044864"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807807.158, "dur": 1.169,
    "args": {
      "External id": 539113,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "9175936"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807818.234, "dur": 1.008,
    "args": {
      "External id": 539114,"Record function id": 0, "Concrete Inputs": ["", "[128, 1024]", "[1024, 1]", "9307008"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807828.712, "dur": 1.069,
    "args": {
      "External id": 539115,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "9438080"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807839.322, "dur": 1.381,
    "args": {
      "External id": 539116,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "9438208"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807850.953, "dur": 1.506,
    "args": {
      "External id": 539117,"Record function id": 0, "Concrete Inputs": ["", "[352, 1024]", "[1024, 1]", "9798656"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807862.238, "dur": 1.092,
    "args": {
      "External id": 539118,"Record function id": 0, "Concrete Inputs": ["", "[128, 2816]", "[2816, 1]", "10159104"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807872.440, "dur": 1.555,
    "args": {
      "External id": 539119,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "10519552"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 31367,
    "ts": 679622807884.025, "dur": 1.253,
    "args": {
      "External id": 539120,"Record function id": 0, "Concrete Inputs": ["", "[4000, 1024]", "[1024, 1]", "10519680"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4591
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#28159", "pid": 4183438, "tid": 4183438,
    "ts": 679622022066.242, "dur": 811969.575,
    "args": {
      "External id": 530433,"Record function id": 0, "Ev Idx": 4592
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.zero_grad#AdamW.zero_grad", "pid": 4183438, "tid": 4183438,
    "ts": 679622022110.335, "dur": 387.024,
    "args": {
      "External id": 530434,"Record function id": 0, "Ev Idx": 4593
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "enumerate(DataLoader)#_StatefulMultiProcessingDataLoaderIter.__next__", "pid": 4183438, "tid": 4183438,
    "ts": 679622022533.631, "dur": 2199.706,
    "args": {
      "External id": 530435,"Record function id": 0, "Ev Idx": 4594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622023621.461, "dur": 8.576,
    "args": {
      "External id": 530436,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 4183438, "tid": 4183438,
    "ts": 679622023682.458, "dur": 7.070,
    "args": {
      "External id": 530437,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "0", "[16, 8192]", "[8192, 1]"], "Input type": ["long int", "", "Scalar", "ScalarList", "ScalarList"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[0], [], [], [], []], "Ev Idx": 4596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622024160.422, "dur": 2.504,
    "args": {
      "External id": 530438,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 4183438, "tid": 4183438,
    "ts": 679622024170.942, "dur": 2.397,
    "args": {
      "External id": 530439,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "0", "[16, 8192]", "[8192, 1]"], "Input type": ["long int", "", "Scalar", "ScalarList", "ScalarList"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[0], [], [], [], []], "Ev Idx": 4598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622024560.544, "dur": 1.380,
    "args": {
      "External id": 530440,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 4183438, "tid": 4183438,
    "ts": 679622024567.774, "dur": 1.813,
    "args": {
      "External id": 530441,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "0", "[16, 8192]", "[8192, 1]"], "Input type": ["long int", "", "Scalar", "ScalarList", "ScalarList"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[0], [], [], [], []], "Ev Idx": 4600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622025456.833, "dur": 14.362,
    "args": {
      "External id": 530442,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], []], "Ev Idx": 4601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622025464.644, "dur": 2.454,
    "args": {
      "External id": 530443,"Record function id": 0, "Concrete Inputs": ["", "[16, 8192]", "[8192, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 4602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622025472.980, "dur": 4.432,
    "args": {
      "External id": 530444,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "4096", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], []], "Ev Idx": 4603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622025475.299, "dur": 1.125,
    "args": {
      "External id": 530445,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 4604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622025503.729, "dur": 554.313,
    "args": {
      "External id": 530446,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], [], [], []], "Ev Idx": 4605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622025511.766, "dur": 545.306,
    "args": {
      "External id": 530447,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], [], []], "Ev Idx": 4606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622025521.137, "dur": 10.042,
    "args": {
      "External id": 530448,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "[4096, 1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622025533.378, "dur": 521.770,
    "args": {
      "External id": 530449,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 4608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622025543.266, "dur": 0.482,
    "args": {
      "External id": 530450,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 4609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622025546.745, "dur": 8.943,
    "args": {
      "External id": 530451,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["long int", "long int"], "Input Strides": [[8192, 1], [4096, 1]], "Input Dims": [[16, 4096], [16, 4096]], "Ev Idx": 4610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 4183438, "tid": 4183438,
    "ts": 679622025552.059, "dur": 3.250,
    "args": {
      "External id": 530452,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "False"], "Input type": ["long int", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], []], "Input Dims": [[16, 4096], [], []], "Ev Idx": 4611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622025554.586, "dur": 0.434,
    "args": {
      "External id": 530453,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 4612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622025558.077, "dur": 260.800,
    "args": {
      "External id": 530454,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 4613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622025560.617, "dur": 257.967,
    "args": {
      "External id": 530455,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 4614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622025562.890, "dur": 15.391,
    "args": {
      "External id": 530456,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 4615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622025565.025, "dur": 12.534,
    "args": {
      "External id": 530457,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622025579.065, "dur": 238.306,
    "args": {
      "External id": 530458,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 4617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622025821.697, "dur": 229.809,
    "args": {
      "External id": 530459,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [4096, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 4618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622026076.966, "dur": 501.409,
    "args": {
      "External id": 530460,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], [], [], [], []], "Ev Idx": 4619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622026079.201, "dur": 498.469,
    "args": {
      "External id": 530461,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], [], [], []], "Ev Idx": 4620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622026086.648, "dur": 8.330,
    "args": {
      "External id": 530462,"Record function id": 0, "Concrete Inputs": ["[16, 8192]", "[8192, 1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622026111.715, "dur": 462.138,
    "args": {
      "External id": 530463,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "Input Dims": [[16, 8192], [16, 8192], []], "Ev Idx": 4622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 4183438, "tid": 4183438,
    "ts": 679622026608.481, "dur": 96.569,
    "args": {
      "External id": 530464,"Record function id": 0, "Concrete Inputs": ["0", "4096", "", "", "", "False"], "Input type": ["Scalar", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622026614.223, "dur": 5.801,
    "args": {
      "External id": 530465,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 4183438, "tid": 4183438,
    "ts": 679622026622.940, "dur": 81.131,
    "args": {
      "External id": 530466,"Record function id": 0, "Concrete Inputs": ["0", "4096", "1", ""], "Input type": ["Scalar", "Scalar", "Scalar", "long int"], "Input Strides": [[], [], [], [1]], "Input Dims": [[], [], [], [0]], "Ev Idx": 4625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622026628.528, "dur": 6.863,
    "args": {
      "External id": 530467,"Record function id": 0, "Concrete Inputs": ["", "[4096]", ""], "Input type": ["long int", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 4626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::repeat", "pid": 4183438, "tid": 4183438,
    "ts": 679622026719.393, "dur": 75.274,
    "args": {
      "External id": 530468,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4096], []], "Ev Idx": 4627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 4183438, "tid": 4183438,
    "ts": 679622026724.220, "dur": 6.808,
    "args": {
      "External id": 530469,"Record function id": 0, "Concrete Inputs": ["", "[1, 4096]", "False"], "Input type": ["long int", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[4096], [], []], "Ev Idx": 4628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622026728.818, "dur": 1.935,
    "args": {
      "External id": 530470,"Record function id": 0, "Concrete Inputs": ["", "[1, 4096]", "[4096, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[4096], [], [], []], "Ev Idx": 4629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622026732.260, "dur": 5.378,
    "args": {
      "External id": 530471,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 4183438,
    "ts": 679622026743.312, "dur": 3.412,
    "args": {
      "External id": 530472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[4096, 1]], "Input Dims": [[16, 4096]], "Ev Idx": 4631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unfold", "pid": 4183438, "tid": 4183438,
    "ts": 679622026749.287, "dur": 5.162,
    "args": {
      "External id": 530473,"Record function id": 0, "Concrete Inputs": ["", "0", "1", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 4632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622026753.708, "dur": 0.572,
    "args": {
      "External id": 530474,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1]", "[4096, 1, 4096]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 4633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unfold", "pid": 4183438, "tid": 4183438,
    "ts": 679622026755.218, "dur": 5.336,
    "args": {
      "External id": 530475,"Record function id": 0, "Concrete Inputs": ["", "1", "4096", "4096"], "Input type": ["long int", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 4096], [], [], []], "Input Dims": [[16, 4096, 1], [], [], []], "Ev Idx": 4634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622026759.541, "dur": 0.921,
    "args": {
      "External id": 530476,"Record function id": 0, "Concrete Inputs": ["", "[16, 1, 1, 4096]", "[4096, 4096, 4096, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1, 4096], [], [], []], "Input Dims": [[16, 4096, 1], [], [], []], "Ev Idx": 4635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622026762.592, "dur": 3.843,
    "args": {
      "External id": 530477,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["long int", "long int"], "Input Strides": [[4096, 1], [4096, 4096, 4096, 1]], "Input Dims": [[1, 4096], [16, 1, 1, 4096]], "Ev Idx": 4636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 4183438, "tid": 4183438,
    "ts": 679622026763.601, "dur": 2.553,
    "args": {
      "External id": 530478,"Record function id": 0, "Concrete Inputs": ["", "[16, 1, 1, 4096]", "False"], "Input type": ["long int", "ScalarList", "Scalar"], "Input Strides": [[4096, 1], [], []], "Input Dims": [[1, 4096], [], []], "Ev Idx": 4637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622026765.124, "dur": 0.934,
    "args": {
      "External id": 530479,"Record function id": 0, "Concrete Inputs": ["", "[16, 1, 1, 4096]", "[0, 4096, 4096, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[1, 4096], [], [], []], "Ev Idx": 4638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622026767.434, "dur": 26.357,
    "args": {
      "External id": 530480,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 4096, 4096, 1], [0, 4096, 4096, 1], []], "Input Dims": [[16, 1, 1, 4096], [16, 1, 1, 4096], []], "Ev Idx": 4639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622026803.467, "dur": 32.584,
    "args": {
      "External id": 530481,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[4096, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 4640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622026805.042, "dur": 30.777,
    "args": {
      "External id": 530482,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[4096, 1], [], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], [], []], "Ev Idx": 4641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622026812.820, "dur": 3.262,
    "args": {
      "External id": 530483,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "[4096, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622026817.143, "dur": 18.086,
    "args": {
      "External id": 530484,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[4096, 1], [4096, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 4643
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::root_pre_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622026944.611, "dur": 146.208,
    "args": {
      "External id": 530485,"Record function id": 0, "Ev Idx": 4644
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::inputs_to_device", "pid": 4183438, "tid": 4183438,
    "ts": 679622027023.401, "dur": 56.360,
    "args": {
      "External id": 530486,"Record function id": 0, "Ev Idx": 4645
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622027114.772, "dur": 45.983,
    "args": {
      "External id": 530487,"Record function id": 0, "Ev Idx": 4646
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622027169.999, "dur": 7791.926,
    "args": {
      "External id": 530488,"Record function id": 0, "Ev Idx": 4647
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather", "pid": 4183438, "tid": 4183438,
    "ts": 679622027179.137, "dur": 1226.424,
    "args": {
      "External id": 530489,"Record function id": 0, "Ev Idx": 4648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622027274.438, "dur": 8.053,
    "args": {
      "External id": 530490,"Record function id": 0, "Concrete Inputs": ["[14615680]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622027300.753, "dur": 126.878,
    "args": {
      "External id": 530491,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["c10::BFloat16", "", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[14615680], [], []], "Ev Idx": 4650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027305.689, "dur": 1.637,
    "args": {
      "External id": 530492,"Record function id": 0, "Concrete Inputs": ["", "[4096000]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027314.181, "dur": 0.537,
    "args": {
      "External id": 530493,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "4096000"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027315.576, "dur": 0.432,
    "args": {
      "External id": 530494,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "4096128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027317.076, "dur": 1.597,
    "args": {
      "External id": 530495,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "4227200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027324.500, "dur": 0.417,
    "args": {
      "External id": 530496,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "4358272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027325.558, "dur": 0.149,
    "args": {
      "External id": 530497,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "4489344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027326.189, "dur": 3.025,
    "args": {
      "External id": 530498,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "4620416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027332.317, "dur": 0.493,
    "args": {
      "External id": 530499,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4620544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027333.503, "dur": 0.369,
    "args": {
      "External id": 530500,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4980992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027338.710, "dur": 0.184,
    "args": {
      "External id": 530501,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "5341440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027339.357, "dur": 0.383,
    "args": {
      "External id": 530502,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "5701888"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027340.467, "dur": 1.409,
    "args": {
      "External id": 530503,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "5702016"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027345.049, "dur": 0.198,
    "args": {
      "External id": 530504,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "5833088"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027346.029, "dur": 0.307,
    "args": {
      "External id": 530505,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "5964160"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027347.109, "dur": 2.474,
    "args": {
      "External id": 530506,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "6095232"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027352.468, "dur": 0.401,
    "args": {
      "External id": 530507,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "6226304"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027353.593, "dur": 0.487,
    "args": {
      "External id": 530508,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "6226432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027358.987, "dur": 0.380,
    "args": {
      "External id": 530509,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "6586880"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027360.106, "dur": 0.379,
    "args": {
      "External id": 530510,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "6947328"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027361.227, "dur": 1.538,
    "args": {
      "External id": 530511,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "7307776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027365.772, "dur": 0.430,
    "args": {
      "External id": 530512,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "7307904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027367.040, "dur": 0.146,
    "args": {
      "External id": 530513,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "7438976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027367.840, "dur": 2.589,
    "args": {
      "External id": 530514,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "7570048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027373.498, "dur": 0.370,
    "args": {
      "External id": 530515,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "7701120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027374.565, "dur": 0.411,
    "args": {
      "External id": 530516,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "7832192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027379.805, "dur": 0.194,
    "args": {
      "External id": 530517,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "7832320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027380.759, "dur": 0.353,
    "args": {
      "External id": 530518,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "8192768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027381.882, "dur": 1.373,
    "args": {
      "External id": 530519,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "8553216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027386.611, "dur": 0.368,
    "args": {
      "External id": 530520,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "8913664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027387.556, "dur": 0.161,
    "args": {
      "External id": 530521,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "8913792"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027388.399, "dur": 3.068,
    "args": {
      "External id": 530522,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "9044864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027394.301, "dur": 0.147,
    "args": {
      "External id": 530523,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "9175936"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027394.950, "dur": 0.414,
    "args": {
      "External id": 530524,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "9307008"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027399.536, "dur": 0.383,
    "args": {
      "External id": 530525,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "9438080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027400.677, "dur": 0.362,
    "args": {
      "External id": 530526,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "9438208"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027401.770, "dur": 1.400,
    "args": {
      "External id": 530527,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "9798656"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027406.232, "dur": 0.415,
    "args": {
      "External id": 530528,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "10159104"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027407.345, "dur": 0.155,
    "args": {
      "External id": 530529,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "10519552"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027407.963, "dur": 2.554,
    "args": {
      "External id": 530530,"Record function id": 0, "Concrete Inputs": ["", "[4096000]", "[1]", "10519680"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622027451.729, "dur": 50.389,
    "args": {
      "External id": 530531,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 4690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622027564.841, "dur": 275.725,
    "args": {
      "External id": 530532,"Record function id": 0, "Concrete Inputs": ["", "", "14615680", "8", "2", "15", ""], "Input type": ["TensorList", "", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 4691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622027578.289, "dur": 4.383,
    "args": {
      "External id": 530533,"Record function id": 0, "Concrete Inputs": ["[116925440]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622027588.828, "dur": 13.579,
    "args": {
      "External id": 530534,"Record function id": 0, "Concrete Inputs": ["", "0", "29231360", "14615680"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[116925440], [], [], []], "Ev Idx": 4693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622027593.445, "dur": 8.480,
    "args": {
      "External id": 530535,"Record function id": 0, "Concrete Inputs": ["", "0", "29231360", "43847040", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[116925440], [], [], [], []], "Ev Idx": 4694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027598.563, "dur": 0.724,
    "args": {
      "External id": 530536,"Record function id": 0, "Concrete Inputs": ["", "[14615680]", "[1]", "29231360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[116925440], [], [], []], "Ev Idx": 4695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622027609.549, "dur": 153.871,
    "args": {
      "External id": 530537,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["c10::BFloat16", "", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[14615680], [], []], "Ev Idx": 4696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027611.360, "dur": 0.447,
    "args": {
      "External id": 530538,"Record function id": 0, "Concrete Inputs": ["", "[4096000]", "[1]", "29231360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027613.253, "dur": 0.387,
    "args": {
      "External id": 530539,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "33327360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027617.387, "dur": 1.412,
    "args": {
      "External id": 530540,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "33327488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027619.697, "dur": 0.558,
    "args": {
      "External id": 530541,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "33458560"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027623.333, "dur": 0.347,
    "args": {
      "External id": 530542,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "33589632"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027624.531, "dur": 0.289,
    "args": {
      "External id": 530543,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "33720704"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027626.017, "dur": 0.512,
    "args": {
      "External id": 530544,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "33851776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027629.039, "dur": 0.674,
    "args": {
      "External id": 530545,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "33851904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027630.454, "dur": 0.692,
    "args": {
      "External id": 530546,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "34212352"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027632.047, "dur": 0.289,
    "args": {
      "External id": 530547,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "34572800"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027635.657, "dur": 1.144,
    "args": {
      "External id": 530548,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "34933248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027637.726, "dur": 0.155,
    "args": {
      "External id": 530549,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "34933376"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027641.131, "dur": 2.072,
    "args": {
      "External id": 530550,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "35064448"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027643.866, "dur": 0.340,
    "args": {
      "External id": 530551,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "35195520"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027644.971, "dur": 0.262,
    "args": {
      "External id": 530552,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "35326592"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027650.430, "dur": 0.165,
    "args": {
      "External id": 530553,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "35457664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027651.346, "dur": 0.258,
    "args": {
      "External id": 530554,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "35457792"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027652.478, "dur": 0.447,
    "args": {
      "External id": 530555,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "35818240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027695.048, "dur": 1.558,
    "args": {
      "External id": 530556,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "36178688"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027697.771, "dur": 0.302,
    "args": {
      "External id": 530557,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "36539136"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027700.945, "dur": 2.152,
    "args": {
      "External id": 530558,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "36539264"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027703.838, "dur": 0.526,
    "args": {
      "External id": 530559,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "36670336"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027705.106, "dur": 0.278,
    "args": {
      "External id": 530560,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "36801408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027710.661, "dur": 0.157,
    "args": {
      "External id": 530561,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "36932480"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027711.569, "dur": 0.395,
    "args": {
      "External id": 530562,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "37063552"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027712.844, "dur": 0.290,
    "args": {
      "External id": 530563,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "37063680"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027718.226, "dur": 1.406,
    "args": {
      "External id": 530564,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "37424128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027720.419, "dur": 0.245,
    "args": {
      "External id": 530565,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "37784576"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027723.651, "dur": 1.972,
    "args": {
      "External id": 530566,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "38145024"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027726.360, "dur": 0.190,
    "args": {
      "External id": 530567,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "38145152"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027727.288, "dur": 0.244,
    "args": {
      "External id": 530568,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "38276224"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027732.195, "dur": 0.215,
    "args": {
      "External id": 530569,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "38407296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027733.151, "dur": 0.295,
    "args": {
      "External id": 530570,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "38538368"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027734.222, "dur": 0.319,
    "args": {
      "External id": 530571,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "38669440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027739.818, "dur": 1.398,
    "args": {
      "External id": 530572,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "38669568"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027742.030, "dur": 0.325,
    "args": {
      "External id": 530573,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "39030016"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027745.512, "dur": 2.004,
    "args": {
      "External id": 530574,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "39390464"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027748.155, "dur": 0.356,
    "args": {
      "External id": 530575,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "39750912"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622027749.255, "dur": 0.360,
    "args": {
      "External id": 530576,"Record function id": 0, "Concrete Inputs": ["", "[4096000]", "[1]", "39751040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[14615680], [], [], []], "Ev Idx": 4735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622027787.600, "dur": 37.470,
    "args": {
      "External id": 530577,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 4736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622027897.366, "dur": 398.126,
    "args": {
      "External id": 530578,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[116925440], [14615680], [], [], []], "Ev Idx": 4737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622027932.941, "dur": 356.820,
    "args": {
      "External id": 530579,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 116925440, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[14615680], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4738, "In msg nelems": 14615680
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622027943.717, "dur": 339.920,
    "args": {
      "External id": 530580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[14615680]], "Ev Idx": 4739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622028319.840, "dur": 2.537,
    "args": {
      "External id": 530581,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4740, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out", "pid": 4183438, "tid": 4183438,
    "ts": 679622028422.808, "dur": 6368.688,
    "args": {
      "External id": 530582,"Record function id": 0, "Ev Idx": 4741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028595.041, "dur": 6.813,
    "args": {
      "External id": 530583,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[116925440], []], "Ev Idx": 4742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028605.616, "dur": 0.987,
    "args": {
      "External id": 530584,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[32768000], []], "Ev Idx": 4743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028608.291, "dur": 0.678,
    "args": {
      "External id": 530585,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028610.623, "dur": 1.991,
    "args": {
      "External id": 530586,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028616.369, "dur": 0.663,
    "args": {
      "External id": 530587,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028618.703, "dur": 0.673,
    "args": {
      "External id": 530588,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028620.891, "dur": 0.854,
    "args": {
      "External id": 530589,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028623.252, "dur": 1.631,
    "args": {
      "External id": 530590,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028630.722, "dur": 0.596,
    "args": {
      "External id": 530591,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028632.804, "dur": 0.619,
    "args": {
      "External id": 530592,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028634.879, "dur": 0.443,
    "args": {
      "External id": 530593,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028636.651, "dur": 1.510,
    "args": {
      "External id": 530594,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028641.983, "dur": 0.699,
    "args": {
      "External id": 530595,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028644.082, "dur": 0.583,
    "args": {
      "External id": 530596,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028645.875, "dur": 0.592,
    "args": {
      "External id": 530597,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028647.878, "dur": 1.905,
    "args": {
      "External id": 530598,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028691.473, "dur": 1.739,
    "args": {
      "External id": 530599,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028696.852, "dur": 1.052,
    "args": {
      "External id": 530600,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028699.330, "dur": 0.539,
    "args": {
      "External id": 530601,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028701.436, "dur": 1.767,
    "args": {
      "External id": 530602,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028706.867, "dur": 0.667,
    "args": {
      "External id": 530603,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028708.847, "dur": 0.621,
    "args": {
      "External id": 530604,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028710.847, "dur": 0.698,
    "args": {
      "External id": 530605,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028712.887, "dur": 1.786,
    "args": {
      "External id": 530606,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028734.869, "dur": 0.795,
    "args": {
      "External id": 530607,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028737.011, "dur": 0.639,
    "args": {
      "External id": 530608,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028739.093, "dur": 0.683,
    "args": {
      "External id": 530609,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028741.407, "dur": 1.315,
    "args": {
      "External id": 530610,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028745.933, "dur": 0.609,
    "args": {
      "External id": 530611,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028747.792, "dur": 0.786,
    "args": {
      "External id": 530612,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028749.774, "dur": 0.795,
    "args": {
      "External id": 530613,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028751.838, "dur": 1.574,
    "args": {
      "External id": 530614,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028758.099, "dur": 0.570,
    "args": {
      "External id": 530615,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028760.279, "dur": 0.559,
    "args": {
      "External id": 530616,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028762.213, "dur": 0.625,
    "args": {
      "External id": 530617,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028764.386, "dur": 1.564,
    "args": {
      "External id": 530618,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028769.125, "dur": 0.487,
    "args": {
      "External id": 530619,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028771.019, "dur": 0.694,
    "args": {
      "External id": 530620,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028773.210, "dur": 0.590,
    "args": {
      "External id": 530621,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622028775.136, "dur": 1.900,
    "args": {
      "External id": 530622,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[32768000], []], "Ev Idx": 4781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622028809.714, "dur": 5927.198,
    "args": {
      "External id": 530623,"Record function id": 0, "Concrete Inputs": ["", "", "1", ""], "Input type": ["c10::BFloat16", "", "Scalar", "TensorList"], "Input Strides": [[14615680, 1], [], [], []], "Input Dims": [[8, 14615680], [], [], []], "Ev Idx": 4782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622028834.163, "dur": 5893.630,
    "args": {
      "External id": 530624,"Record function id": 0, "Concrete Inputs": ["", "", "1", ""], "Input type": ["c10::BFloat16", "", "Scalar", "TensorList"], "Input Strides": [[14615680, 1], [], [], []], "Input Dims": [[8, 14615680], [], [], []], "Ev Idx": 4783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622028854.620, "dur": 5.650,
    "args": {
      "External id": 530625,"Record function id": 0, "Concrete Inputs": ["[3034]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622028864.804, "dur": 5818.506,
    "args": {
      "External id": 530626,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[3034], [], [], [], [], [], [], []], "Ev Idx": 4785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622028867.413, "dur": 5815.261,
    "args": {
      "External id": 530627,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[3034], [], [], [], [], [], []], "Ev Idx": 4786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622028875.639, "dur": 6.513,
    "args": {
      "External id": 530628,"Record function id": 0, "Concrete Inputs": ["[3034]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622028883.638, "dur": 5795.750,
    "args": {
      "External id": 530629,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[3034], [3034], []], "Ev Idx": 4788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622035040.185, "dur": 35.618,
    "args": {
      "External id": 530630,"Record function id": 0, "Ev Idx": 4789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 4183438, "tid": 4183438,
    "ts": 679622035077.078, "dur": 222.773,
    "args": {
      "External id": 530631,"Record function id": 0, "Ev Idx": 4790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622035132.268, "dur": 158.431,
    "args": {
      "External id": 530632,"Sequence number": 6319416, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "long int"], "Input Strides": [[1024, 1], [4096, 1]], "Input Dims": [[32000, 1024], [16, 4096]], "Ev Idx": 4791
    }
  },
  {
    "ph": "s", "id": 212, "pid": 4183438, "tid": 4183438, "ts": 679622035132.268,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622035206.540, "dur": 47.560,
    "args": {
      "External id": 530633,"kernel_hash": "crydi4fszmveuo3xtgmvd5wzmvtv7cilsx6l2c4ussrew7shlvlg", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/ry/crydi4fszmveuo3xtgmvd5wzmvtv7cilsx6l2c4ussrew7shlvlg.py", "kernel_backend": "triton", "Input type": ["long int", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], [1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096], [32000, 1024], [16, 4096, 1024], []], "Ev Idx": 4792
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622035364.404, "dur": 55.462,
    "args": {
      "External id": 530634,"Record function id": 0, "Ev Idx": 4793
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.0)", "pid": 4183438, "tid": 4183438,
    "ts": 679622035431.275, "dur": 6828.629,
    "args": {
      "External id": 530635,"Record function id": 0, "Ev Idx": 4794
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.0)", "pid": 4183438, "tid": 4183438,
    "ts": 679622035440.576, "dur": 836.340,
    "args": {
      "External id": 530636,"Record function id": 0, "Ev Idx": 4795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622035508.066, "dur": 11.525,
    "args": {
      "External id": 530637,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622035532.608, "dur": 44.048,
    "args": {
      "External id": 530638,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035542.633, "dur": 2.274,
    "args": {
      "External id": 530639,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035546.674, "dur": 0.452,
    "args": {
      "External id": 530640,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035547.843, "dur": 2.157,
    "args": {
      "External id": 530641,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035553.358, "dur": 0.305,
    "args": {
      "External id": 530642,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035554.341, "dur": 0.285,
    "args": {
      "External id": 530643,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035557.779, "dur": 0.180,
    "args": {
      "External id": 530644,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035561.603, "dur": 0.305,
    "args": {
      "External id": 530645,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035562.776, "dur": 1.403,
    "args": {
      "External id": 530646,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035569.684, "dur": 0.332,
    "args": {
      "External id": 530647,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622035587.345, "dur": 35.205,
    "args": {
      "External id": 530648,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622035696.778, "dur": 122.088,
    "args": {
      "External id": 530649,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 4808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622035709.113, "dur": 5.752,
    "args": {
      "External id": 530650,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622035720.767, "dur": 13.042,
    "args": {
      "External id": 530651,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622035725.166, "dur": 8.233,
    "args": {
      "External id": 530652,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 4811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035731.051, "dur": 0.764,
    "args": {
      "External id": 530653,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622035741.291, "dur": 31.690,
    "args": {
      "External id": 530654,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035743.262, "dur": 0.427,
    "args": {
      "External id": 530655,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035746.931, "dur": 0.570,
    "args": {
      "External id": 530656,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035751.080, "dur": 0.191,
    "args": {
      "External id": 530657,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035752.228, "dur": 0.312,
    "args": {
      "External id": 530658,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035755.651, "dur": 3.696,
    "args": {
      "External id": 530659,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035760.515, "dur": 0.190,
    "args": {
      "External id": 530660,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035764.625, "dur": 0.152,
    "args": {
      "External id": 530661,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035767.715, "dur": 0.319,
    "args": {
      "External id": 530662,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622035768.604, "dur": 0.338,
    "args": {
      "External id": 530663,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622035786.365, "dur": 24.197,
    "args": {
      "External id": 530664,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622035870.129, "dur": 313.253,
    "args": {
      "External id": 530665,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 4824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622035900.166, "dur": 278.305,
    "args": {
      "External id": 530666,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4825, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622035911.774, "dur": 260.680,
    "args": {
      "External id": 530667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 4826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622036207.834, "dur": 2.300,
    "args": {
      "External id": 530668,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4827, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.0)", "pid": 4183438, "tid": 4183438,
    "ts": 679622036298.021, "dur": 5683.395,
    "args": {
      "External id": 530669,"Record function id": 0, "Ev Idx": 4828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036389.815, "dur": 6.567,
    "args": {
      "External id": 530670,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036399.687, "dur": 1.019,
    "args": {
      "External id": 530671,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036402.787, "dur": 0.993,
    "args": {
      "External id": 530672,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036405.546, "dur": 1.937,
    "args": {
      "External id": 530673,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036408.904, "dur": 0.793,
    "args": {
      "External id": 530674,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036413.387, "dur": 0.716,
    "args": {
      "External id": 530675,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036416.001, "dur": 0.664,
    "args": {
      "External id": 530676,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036418.113, "dur": 2.192,
    "args": {
      "External id": 530677,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036424.024, "dur": 0.592,
    "args": {
      "External id": 530678,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622036428.032, "dur": 0.753,
    "args": {
      "External id": 530679,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622036445.500, "dur": 5483.550,
    "args": {
      "External id": 530680,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622036462.625, "dur": 5457.893,
    "args": {
      "External id": 530681,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622036486.245, "dur": 14.078,
    "args": {
      "External id": 530682,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622036502.967, "dur": 5376.658,
    "args": {
      "External id": 530683,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 4842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622036505.468, "dur": 5373.603,
    "args": {
      "External id": 530684,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 4843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622036511.420, "dur": 7.022,
    "args": {
      "External id": 530685,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622036520.200, "dur": 5356.041,
    "args": {
      "External id": 530686,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 4845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622042182.624, "dur": 42.316,
    "args": {
      "External id": 530687,"Sequence number": 6319417, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4846
    }
  },
  {
    "ph": "s", "id": 211, "pid": 4183438, "tid": 4183438, "ts": 679622042182.624,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622042203.137, "dur": 15.832,
    "args": {
      "External id": 530688,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 4847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622042211.392, "dur": 7.005,
    "args": {
      "External id": 530689,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 4848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622042309.494, "dur": 106.922,
    "args": {
      "External id": 530690,"Record function id": 0, "Ev Idx": 4849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622042418.927, "dur": 1769.628,
    "args": {
      "External id": 530691,"Record function id": 0, "Ev Idx": 4850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622042475.568, "dur": 1691.386,
    "args": {
      "External id": 530692,"Sequence number": 6319418, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 4851
    }
  },
  {
    "ph": "s", "id": 210, "pid": 4183438, "tid": 4183438, "ts": 679622042475.568,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622042559.636, "dur": 59.418,
    "args": {
      "External id": 530693,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 4852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622042636.393, "dur": 167.861,
    "args": {
      "External id": 530694,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622042824.993, "dur": 53.122,
    "args": {
      "External id": 530695,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622042891.413, "dur": 39.774,
    "args": {
      "External id": 530696,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622042968.704, "dur": 36.161,
    "args": {
      "External id": 530697,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622043031.772, "dur": 20.710,
    "args": {
      "External id": 530698,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622043078.229, "dur": 223.923,
    "args": {
      "External id": 530699,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 4858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622043186.000, "dur": 19.480,
    "args": {
      "External id": 530700,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 4859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622043192.569, "dur": 11.152,
    "args": {
      "External id": 530701,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622043209.081, "dur": 4.518,
    "args": {
      "External id": 530702,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622043217.818, "dur": 1.901,
    "args": {
      "External id": 530703,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622043222.574, "dur": 3.562,
    "args": {
      "External id": 530704,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622043318.440, "dur": 113.354,
    "args": {
      "External id": 530705,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622043539.191, "dur": 47.701,
    "args": {
      "External id": 530706,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 4865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622043600.203, "dur": 143.030,
    "args": {
      "External id": 530707,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622043768.824, "dur": 63.338,
    "args": {
      "External id": 530708,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622043871.547, "dur": 37.118,
    "args": {
      "External id": 530709,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 4868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622043917.868, "dur": 48.939,
    "args": {
      "External id": 530710,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622044000.863, "dur": 29.904,
    "args": {
      "External id": 530711,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 4870
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.0)", "pid": 4183438, "tid": 4183438,
    "ts": 679622044271.860, "dur": 99.305,
    "args": {
      "External id": 530712,"Record function id": 0, "Ev Idx": 4871
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622044463.711, "dur": 66.494,
    "args": {
      "External id": 530713,"Record function id": 0, "Ev Idx": 4872
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.1)", "pid": 4183438, "tid": 4183438,
    "ts": 679622044542.803, "dur": 8485.627,
    "args": {
      "External id": 530714,"Record function id": 0, "Ev Idx": 4873
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.1)", "pid": 4183438, "tid": 4183438,
    "ts": 679622044556.730, "dur": 1353.967,
    "args": {
      "External id": 530715,"Record function id": 0, "Ev Idx": 4874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622044700.950, "dur": 13.928,
    "args": {
      "External id": 530716,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622044733.121, "dur": 56.026,
    "args": {
      "External id": 530717,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044742.119, "dur": 5.930,
    "args": {
      "External id": 530718,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044750.635, "dur": 0.447,
    "args": {
      "External id": 530719,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044752.748, "dur": 0.596,
    "args": {
      "External id": 530720,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044759.725, "dur": 0.416,
    "args": {
      "External id": 530721,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044761.692, "dur": 1.866,
    "args": {
      "External id": 530722,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044767.320, "dur": 0.411,
    "args": {
      "External id": 530723,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044771.869, "dur": 0.432,
    "args": {
      "External id": 530724,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044773.477, "dur": 0.319,
    "args": {
      "External id": 530725,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044778.269, "dur": 2.980,
    "args": {
      "External id": 530726,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622044802.216, "dur": 54.241,
    "args": {
      "External id": 530727,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622044903.916, "dur": 145.946,
    "args": {
      "External id": 530728,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 4887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622044917.377, "dur": 5.009,
    "args": {
      "External id": 530729,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622044929.479, "dur": 14.536,
    "args": {
      "External id": 530730,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622044934.631, "dur": 8.790,
    "args": {
      "External id": 530731,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 4890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044941.258, "dur": 0.645,
    "args": {
      "External id": 530732,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622044951.813, "dur": 36.216,
    "args": {
      "External id": 530733,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044953.940, "dur": 0.488,
    "args": {
      "External id": 530734,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044958.320, "dur": 1.647,
    "args": {
      "External id": 530735,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044961.293, "dur": 2.995,
    "args": {
      "External id": 530736,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044967.333, "dur": 0.414,
    "args": {
      "External id": 530737,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044969.251, "dur": 0.262,
    "args": {
      "External id": 530738,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044973.636, "dur": 0.300,
    "args": {
      "External id": 530739,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044977.706, "dur": 0.271,
    "args": {
      "External id": 530740,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044979.097, "dur": 0.276,
    "args": {
      "External id": 530741,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622044982.629, "dur": 0.244,
    "args": {
      "External id": 530742,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622045006.349, "dur": 31.057,
    "args": {
      "External id": 530743,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622045157.119, "dur": 611.210,
    "args": {
      "External id": 530744,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 4903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622045194.139, "dur": 564.805,
    "args": {
      "External id": 530745,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4904, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622045209.848, "dur": 537.901,
    "args": {
      "External id": 530746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 4905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622045804.559, "dur": 4.055,
    "args": {
      "External id": 530747,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4906, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.1)", "pid": 4183438, "tid": 4183438,
    "ts": 679622045939.287, "dur": 6820.440,
    "args": {
      "External id": 530748,"Record function id": 0, "Ev Idx": 4907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046066.495, "dur": 8.101,
    "args": {
      "External id": 530749,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046079.186, "dur": 1.436,
    "args": {
      "External id": 530750,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046083.108, "dur": 1.158,
    "args": {
      "External id": 530751,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046087.262, "dur": 1.114,
    "args": {
      "External id": 530752,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046090.328, "dur": 1.220,
    "args": {
      "External id": 530753,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046095.505, "dur": 32.316,
    "args": {
      "External id": 530754,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046132.909, "dur": 1.585,
    "args": {
      "External id": 530755,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046139.272, "dur": 3.556,
    "args": {
      "External id": 530756,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046145.074, "dur": 1.017,
    "args": {
      "External id": 530757,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622046149.894, "dur": 0.794,
    "args": {
      "External id": 530758,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622046174.942, "dur": 6512.840,
    "args": {
      "External id": 530759,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622046196.831, "dur": 6445.878,
    "args": {
      "External id": 530760,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622046227.906, "dur": 18.116,
    "args": {
      "External id": 530761,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622046249.491, "dur": 6334.025,
    "args": {
      "External id": 530762,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 4921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622046252.472, "dur": 6329.675,
    "args": {
      "External id": 530763,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 4922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622046259.327, "dur": 6.675,
    "args": {
      "External id": 530764,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622046268.216, "dur": 6307.768,
    "args": {
      "External id": 530765,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 4924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622052947.653, "dur": 42.721,
    "args": {
      "External id": 530766,"Sequence number": 6319419, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 4925
    }
  },
  {
    "ph": "s", "id": 209, "pid": 4183438, "tid": 4183438, "ts": 679622052947.653,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622052968.779, "dur": 15.875,
    "args": {
      "External id": 530767,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 4926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622052976.370, "dur": 7.791,
    "args": {
      "External id": 530768,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 4927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622053076.961, "dur": 131.260,
    "args": {
      "External id": 530769,"Record function id": 0, "Ev Idx": 4928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622053213.034, "dur": 1598.134,
    "args": {
      "External id": 530770,"Record function id": 0, "Ev Idx": 4929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622053270.761, "dur": 1521.904,
    "args": {
      "External id": 530771,"Sequence number": 6319420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 4930
    }
  },
  {
    "ph": "s", "id": 208, "pid": 4183438, "tid": 4183438, "ts": 679622053270.761,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622053374.131, "dur": 59.389,
    "args": {
      "External id": 530772,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 4931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622053454.360, "dur": 111.119,
    "args": {
      "External id": 530773,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622053576.713, "dur": 46.021,
    "args": {
      "External id": 530774,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622053631.201, "dur": 85.040,
    "args": {
      "External id": 530775,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622053774.442, "dur": 39.609,
    "args": {
      "External id": 530776,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622053835.474, "dur": 25.484,
    "args": {
      "External id": 530777,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622053884.992, "dur": 192.245,
    "args": {
      "External id": 530778,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 4937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622053954.937, "dur": 20.526,
    "args": {
      "External id": 530779,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 4938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622053961.577, "dur": 12.656,
    "args": {
      "External id": 530780,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622053980.692, "dur": 5.618,
    "args": {
      "External id": 530781,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622053988.271, "dur": 3.959,
    "args": {
      "External id": 530782,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622053995.206, "dur": 3.712,
    "args": {
      "External id": 530783,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622054095.643, "dur": 130.243,
    "args": {
      "External id": 530784,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 4943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622054296.531, "dur": 50.748,
    "args": {
      "External id": 530785,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 4944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622054365.013, "dur": 76.739,
    "args": {
      "External id": 530786,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622054453.211, "dur": 49.591,
    "args": {
      "External id": 530787,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 4946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622054538.307, "dur": 33.914,
    "args": {
      "External id": 530788,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 4947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622054581.646, "dur": 42.680,
    "args": {
      "External id": 530789,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 4948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622054648.532, "dur": 63.472,
    "args": {
      "External id": 530790,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 4949
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.1)", "pid": 4183438, "tid": 4183438,
    "ts": 679622054884.313, "dur": 83.054,
    "args": {
      "External id": 530791,"Record function id": 0, "Ev Idx": 4950
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622055042.510, "dur": 47.146,
    "args": {
      "External id": 530792,"Record function id": 0, "Ev Idx": 4951
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.2)", "pid": 4183438, "tid": 4183438,
    "ts": 679622055121.733, "dur": 8709.376,
    "args": {
      "External id": 530793,"Record function id": 0, "Ev Idx": 4952
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.2)", "pid": 4183438, "tid": 4183438,
    "ts": 679622055136.897, "dur": 901.196,
    "args": {
      "External id": 530794,"Record function id": 0, "Ev Idx": 4953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622055217.848, "dur": 10.702,
    "args": {
      "External id": 530795,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622055241.891, "dur": 41.654,
    "args": {
      "External id": 530796,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055249.261, "dur": 2.415,
    "args": {
      "External id": 530797,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055255.945, "dur": 0.273,
    "args": {
      "External id": 530798,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055259.458, "dur": 0.401,
    "args": {
      "External id": 530799,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055260.707, "dur": 0.357,
    "args": {
      "External id": 530800,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055263.518, "dur": 0.525,
    "args": {
      "External id": 530801,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055266.943, "dur": 0.314,
    "args": {
      "External id": 530802,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055267.891, "dur": 4.554,
    "args": {
      "External id": 530803,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055273.314, "dur": 0.412,
    "args": {
      "External id": 530804,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055277.080, "dur": 0.338,
    "args": {
      "External id": 530805,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622055294.969, "dur": 42.078,
    "args": {
      "External id": 530806,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622055371.395, "dur": 110.702,
    "args": {
      "External id": 530807,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 4966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622055385.783, "dur": 3.792,
    "args": {
      "External id": 530808,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622055395.040, "dur": 9.490,
    "args": {
      "External id": 530809,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622055399.656, "dur": 4.495,
    "args": {
      "External id": 530810,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 4969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055402.532, "dur": 0.469,
    "args": {
      "External id": 530811,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 4970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622055410.366, "dur": 31.681,
    "args": {
      "External id": 530812,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 4971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055414.140, "dur": 2.458,
    "args": {
      "External id": 530813,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055417.367, "dur": 0.295,
    "args": {
      "External id": 530814,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055418.373, "dur": 0.261,
    "args": {
      "External id": 530815,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055424.034, "dur": 1.345,
    "args": {
      "External id": 530816,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055425.987, "dur": 0.154,
    "args": {
      "External id": 530817,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055426.905, "dur": 0.148,
    "args": {
      "External id": 530818,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055430.663, "dur": 0.148,
    "args": {
      "External id": 530819,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055431.664, "dur": 0.312,
    "args": {
      "External id": 530820,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622055435.468, "dur": 2.872,
    "args": {
      "External id": 530821,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 4980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622055451.960, "dur": 22.111,
    "args": {
      "External id": 530822,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 4981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622055534.686, "dur": 406.114,
    "args": {
      "External id": 530823,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 4982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622055568.897, "dur": 366.197,
    "args": {
      "External id": 530824,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4983, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622055578.970, "dur": 349.992,
    "args": {
      "External id": 530825,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 4984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622055964.628, "dur": 2.858,
    "args": {
      "External id": 530826,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4985, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.2)", "pid": 4183438, "tid": 4183438,
    "ts": 679622056058.153, "dur": 7515.887,
    "args": {
      "External id": 530827,"Record function id": 0, "Ev Idx": 4986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056177.589, "dur": 6.947,
    "args": {
      "External id": 530828,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 4987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056188.267, "dur": 1.112,
    "args": {
      "External id": 530829,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056191.131, "dur": 2.144,
    "args": {
      "External id": 530830,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056195.055, "dur": 1.024,
    "args": {
      "External id": 530831,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056197.466, "dur": 0.783,
    "args": {
      "External id": 530832,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056201.647, "dur": 0.755,
    "args": {
      "External id": 530833,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 4992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056205.884, "dur": 0.781,
    "args": {
      "External id": 530834,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 4993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056207.794, "dur": 2.036,
    "args": {
      "External id": 530835,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056211.315, "dur": 0.828,
    "args": {
      "External id": 530836,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622056215.775, "dur": 0.582,
    "args": {
      "External id": 530837,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 4996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622056235.071, "dur": 7264.179,
    "args": {
      "External id": 530838,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622056252.970, "dur": 7230.890,
    "args": {
      "External id": 530839,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 4998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622056270.708, "dur": 14.187,
    "args": {
      "External id": 530840,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622056287.503, "dur": 7136.980,
    "args": {
      "External id": 530841,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622056290.304, "dur": 7132.890,
    "args": {
      "External id": 530842,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622056296.133, "dur": 4.955,
    "args": {
      "External id": 530843,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622056302.848, "dur": 7113.399,
    "args": {
      "External id": 530844,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622063767.010, "dur": 35.841,
    "args": {
      "External id": 530845,"Sequence number": 6319421, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5004
    }
  },
  {
    "ph": "s", "id": 207, "pid": 4183438, "tid": 4183438, "ts": 679622063767.010,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622063784.594, "dur": 13.334,
    "args": {
      "External id": 530846,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622063790.460, "dur": 7.066,
    "args": {
      "External id": 530847,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622063870.573, "dur": 87.818,
    "args": {
      "External id": 530848,"Record function id": 0, "Ev Idx": 5007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622063959.837, "dur": 1074.098,
    "args": {
      "External id": 530849,"Record function id": 0, "Ev Idx": 5008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622063999.082, "dur": 1020.298,
    "args": {
      "External id": 530850,"Sequence number": 6319422, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5009
    }
  },
  {
    "ph": "s", "id": 206, "pid": 4183438, "tid": 4183438, "ts": 679622063999.082,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622064070.566, "dur": 61.020,
    "args": {
      "External id": 530851,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622064149.022, "dur": 93.750,
    "args": {
      "External id": 530852,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622064253.598, "dur": 35.426,
    "args": {
      "External id": 530853,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622064298.023, "dur": 30.887,
    "args": {
      "External id": 530854,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622064354.901, "dur": 26.534,
    "args": {
      "External id": 530855,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622064401.714, "dur": 14.246,
    "args": {
      "External id": 530856,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622064433.825, "dur": 127.928,
    "args": {
      "External id": 530857,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622064483.806, "dur": 12.658,
    "args": {
      "External id": 530858,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622064488.784, "dur": 6.749,
    "args": {
      "External id": 530859,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622064501.446, "dur": 3.707,
    "args": {
      "External id": 530860,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622064506.552, "dur": 1.063,
    "args": {
      "External id": 530861,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622064509.891, "dur": 3.380,
    "args": {
      "External id": 530862,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622064572.502, "dur": 42.222,
    "args": {
      "External id": 530863,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622064642.908, "dur": 66.762,
    "args": {
      "External id": 530864,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622064721.438, "dur": 43.279,
    "args": {
      "External id": 530865,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622064773.696, "dur": 32.349,
    "args": {
      "External id": 530866,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622064830.098, "dur": 24.379,
    "args": {
      "External id": 530867,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622064859.798, "dur": 42.393,
    "args": {
      "External id": 530868,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622064923.629, "dur": 19.866,
    "args": {
      "External id": 530869,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5028
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.2)", "pid": 4183438, "tid": 4183438,
    "ts": 679622065116.044, "dur": 79.150,
    "args": {
      "External id": 530870,"Record function id": 0, "Ev Idx": 5029
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622065270.871, "dur": 48.658,
    "args": {
      "External id": 530871,"Record function id": 0, "Ev Idx": 5030
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.3)", "pid": 4183438, "tid": 4183438,
    "ts": 679622065328.208, "dur": 8603.920,
    "args": {
      "External id": 530872,"Record function id": 0, "Ev Idx": 5031
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.3)", "pid": 4183438, "tid": 4183438,
    "ts": 679622065340.065, "dur": 957.313,
    "args": {
      "External id": 530873,"Record function id": 0, "Ev Idx": 5032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622065420.247, "dur": 8.871,
    "args": {
      "External id": 530874,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622065441.877, "dur": 44.894,
    "args": {
      "External id": 530875,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065449.430, "dur": 2.336,
    "args": {
      "External id": 530876,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065456.194, "dur": 0.414,
    "args": {
      "External id": 530877,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065459.078, "dur": 0.307,
    "args": {
      "External id": 530878,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065465.641, "dur": 0.543,
    "args": {
      "External id": 530879,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065469.349, "dur": 0.433,
    "args": {
      "External id": 530880,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065472.317, "dur": 0.289,
    "args": {
      "External id": 530881,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065473.181, "dur": 3.405,
    "args": {
      "External id": 530882,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065477.090, "dur": 0.193,
    "args": {
      "External id": 530883,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065479.802, "dur": 0.154,
    "args": {
      "External id": 530884,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622065497.148, "dur": 44.784,
    "args": {
      "External id": 530885,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622065575.106, "dur": 177.713,
    "args": {
      "External id": 530886,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622065589.830, "dur": 3.589,
    "args": {
      "External id": 530887,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622065598.536, "dur": 9.576,
    "args": {
      "External id": 530888,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622065603.061, "dur": 4.678,
    "args": {
      "External id": 530889,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065606.059, "dur": 0.424,
    "args": {
      "External id": 530890,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622065614.096, "dur": 34.069,
    "args": {
      "External id": 530891,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065618.316, "dur": 2.749,
    "args": {
      "External id": 530892,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065621.940, "dur": 0.431,
    "args": {
      "External id": 530893,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065623.158, "dur": 0.411,
    "args": {
      "External id": 530894,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065628.852, "dur": 1.354,
    "args": {
      "External id": 530895,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065630.925, "dur": 0.307,
    "args": {
      "External id": 530896,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065631.871, "dur": 0.169,
    "args": {
      "External id": 530897,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065637.218, "dur": 0.152,
    "args": {
      "External id": 530898,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065638.667, "dur": 0.160,
    "args": {
      "External id": 530899,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622065641.015, "dur": 2.163,
    "args": {
      "External id": 530900,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622065713.248, "dur": 29.850,
    "args": {
      "External id": 530901,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622065808.105, "dur": 393.347,
    "args": {
      "External id": 530902,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622065841.251, "dur": 354.947,
    "args": {
      "External id": 530903,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5062, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622065854.099, "dur": 335.905,
    "args": {
      "External id": 530904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622066224.698, "dur": 2.686,
    "args": {
      "External id": 530905,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5064, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.3)", "pid": 4183438, "tid": 4183438,
    "ts": 679622066318.395, "dur": 7420.063,
    "args": {
      "External id": 530906,"Record function id": 0, "Ev Idx": 5065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066410.605, "dur": 6.794,
    "args": {
      "External id": 530907,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066421.178, "dur": 0.942,
    "args": {
      "External id": 530908,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066423.728, "dur": 2.138,
    "args": {
      "External id": 530909,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066427.492, "dur": 0.927,
    "args": {
      "External id": 530910,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066430.021, "dur": 0.790,
    "args": {
      "External id": 530911,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066434.669, "dur": 0.696,
    "args": {
      "External id": 530912,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066438.908, "dur": 0.852,
    "args": {
      "External id": 530913,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066441.245, "dur": 2.153,
    "args": {
      "External id": 530914,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066444.647, "dur": 0.749,
    "args": {
      "External id": 530915,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622066449.199, "dur": 0.556,
    "args": {
      "External id": 530916,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622066466.705, "dur": 7232.739,
    "args": {
      "External id": 530917,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622066484.137, "dur": 7207.544,
    "args": {
      "External id": 530918,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622066501.386, "dur": 14.518,
    "args": {
      "External id": 530919,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622066518.678, "dur": 7116.372,
    "args": {
      "External id": 530920,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622066521.034, "dur": 7113.542,
    "args": {
      "External id": 530921,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622066527.177, "dur": 5.854,
    "args": {
      "External id": 530922,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622066534.708, "dur": 7096.882,
    "args": {
      "External id": 530923,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622073868.954, "dur": 38.902,
    "args": {
      "External id": 530924,"Sequence number": 6319423, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5083
    }
  },
  {
    "ph": "s", "id": 205, "pid": 4183438, "tid": 4183438, "ts": 679622073868.954,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622073893.450, "dur": 9.839,
    "args": {
      "External id": 530925,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622073898.577, "dur": 4.404,
    "args": {
      "External id": 530926,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622073969.505, "dur": 87.277,
    "args": {
      "External id": 530927,"Record function id": 0, "Ev Idx": 5086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622074058.367, "dur": 1145.854,
    "args": {
      "External id": 530928,"Record function id": 0, "Ev Idx": 5087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622074113.805, "dur": 1075.474,
    "args": {
      "External id": 530929,"Sequence number": 6319424, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5088
    }
  },
  {
    "ph": "s", "id": 204, "pid": 4183438, "tid": 4183438, "ts": 679622074113.805,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622074184.159, "dur": 40.693,
    "args": {
      "External id": 530930,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622074237.173, "dur": 95.557,
    "args": {
      "External id": 530931,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622074345.864, "dur": 35.222,
    "args": {
      "External id": 530932,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622074386.531, "dur": 29.409,
    "args": {
      "External id": 530933,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622074446.106, "dur": 26.295,
    "args": {
      "External id": 530934,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622074488.046, "dur": 16.515,
    "args": {
      "External id": 530935,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622074522.055, "dur": 169.831,
    "args": {
      "External id": 530936,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622074572.385, "dur": 10.195,
    "args": {
      "External id": 530937,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622074576.843, "dur": 5.011,
    "args": {
      "External id": 530938,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622074586.734, "dur": 5.465,
    "args": {
      "External id": 530939,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622074593.707, "dur": 1.005,
    "args": {
      "External id": 530940,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622074598.866, "dur": 3.005,
    "args": {
      "External id": 530941,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622074706.213, "dur": 54.662,
    "args": {
      "External id": 530942,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622074793.220, "dur": 28.696,
    "args": {
      "External id": 530943,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622074830.169, "dur": 38.981,
    "args": {
      "External id": 530944,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622074902.736, "dur": 48.053,
    "args": {
      "External id": 530945,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622074978.332, "dur": 26.092,
    "args": {
      "External id": 530946,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622075010.429, "dur": 50.104,
    "args": {
      "External id": 530947,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622075077.241, "dur": 40.527,
    "args": {
      "External id": 530948,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5107
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.3)", "pid": 4183438, "tid": 4183438,
    "ts": 679622075272.721, "dur": 75.505,
    "args": {
      "External id": 530949,"Record function id": 0, "Ev Idx": 5108
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622075419.952, "dur": 48.748,
    "args": {
      "External id": 530950,"Record function id": 0, "Ev Idx": 5109
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.4)", "pid": 4183438, "tid": 4183438,
    "ts": 679622075477.011, "dur": 8888.397,
    "args": {
      "External id": 530951,"Record function id": 0, "Ev Idx": 5110
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.4)", "pid": 4183438, "tid": 4183438,
    "ts": 679622075489.509, "dur": 933.600,
    "args": {
      "External id": 530952,"Record function id": 0, "Ev Idx": 5111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622075566.955, "dur": 9.172,
    "args": {
      "External id": 530953,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622075588.752, "dur": 37.562,
    "args": {
      "External id": 530954,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075594.903, "dur": 2.069,
    "args": {
      "External id": 530955,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075600.996, "dur": 0.623,
    "args": {
      "External id": 530956,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075603.987, "dur": 0.198,
    "args": {
      "External id": 530957,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075605.048, "dur": 0.389,
    "args": {
      "External id": 530958,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075608.825, "dur": 0.431,
    "args": {
      "External id": 530959,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075611.524, "dur": 0.332,
    "args": {
      "External id": 530960,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075612.892, "dur": 3.540,
    "args": {
      "External id": 530961,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075617.321, "dur": 0.273,
    "args": {
      "External id": 530962,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075619.853, "dur": 0.184,
    "args": {
      "External id": 530963,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622075637.020, "dur": 82.453,
    "args": {
      "External id": 530964,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622075758.399, "dur": 118.535,
    "args": {
      "External id": 530965,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622075773.622, "dur": 5.182,
    "args": {
      "External id": 530966,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622075784.790, "dur": 10.487,
    "args": {
      "External id": 530967,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622075789.375, "dur": 5.509,
    "args": {
      "External id": 530968,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075792.649, "dur": 0.638,
    "args": {
      "External id": 530969,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622075801.892, "dur": 32.192,
    "args": {
      "External id": 530970,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075805.264, "dur": 2.247,
    "args": {
      "External id": 530971,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075808.566, "dur": 0.324,
    "args": {
      "External id": 530972,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075809.796, "dur": 0.394,
    "args": {
      "External id": 530973,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075814.947, "dur": 1.472,
    "args": {
      "External id": 530974,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075817.309, "dur": 0.399,
    "args": {
      "External id": 530975,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075818.313, "dur": 0.177,
    "args": {
      "External id": 530976,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075823.897, "dur": 0.191,
    "args": {
      "External id": 530977,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075824.951, "dur": 0.398,
    "args": {
      "External id": 530978,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622075828.038, "dur": 2.498,
    "args": {
      "External id": 530979,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622075845.679, "dur": 23.076,
    "args": {
      "External id": 530980,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622075931.994, "dur": 393.927,
    "args": {
      "External id": 530981,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622075966.737, "dur": 353.466,
    "args": {
      "External id": 530982,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5141, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622075977.015, "dur": 337.379,
    "args": {
      "External id": 530983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622076348.283, "dur": 2.436,
    "args": {
      "External id": 530984,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5143, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.4)", "pid": 4183438, "tid": 4183438,
    "ts": 679622076444.328, "dur": 7629.188,
    "args": {
      "External id": 530985,"Record function id": 0, "Ev Idx": 5144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076532.009, "dur": 6.663,
    "args": {
      "External id": 530986,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076541.981, "dur": 1.157,
    "args": {
      "External id": 530987,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076545.027, "dur": 1.958,
    "args": {
      "External id": 530988,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076548.463, "dur": 0.826,
    "args": {
      "External id": 530989,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076550.884, "dur": 0.760,
    "args": {
      "External id": 530990,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076552.958, "dur": 0.606,
    "args": {
      "External id": 530991,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076558.451, "dur": 0.954,
    "args": {
      "External id": 530992,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076560.854, "dur": 1.883,
    "args": {
      "External id": 530993,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076563.856, "dur": 0.888,
    "args": {
      "External id": 530994,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622076565.898, "dur": 0.847,
    "args": {
      "External id": 530995,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622076585.911, "dur": 7408.323,
    "args": {
      "External id": 530996,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622076602.303, "dur": 7377.559,
    "args": {
      "External id": 530997,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622076619.652, "dur": 14.947,
    "args": {
      "External id": 530998,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622076637.502, "dur": 7284.585,
    "args": {
      "External id": 530999,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622076639.716, "dur": 7281.094,
    "args": {
      "External id": 531000,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622076645.732, "dur": 4.467,
    "args": {
      "External id": 531001,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622076651.990, "dur": 7262.137,
    "args": {
      "External id": 531002,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622084284.487, "dur": 42.481,
    "args": {
      "External id": 531003,"Sequence number": 6319425, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5162
    }
  },
  {
    "ph": "s", "id": 203, "pid": 4183438, "tid": 4183438, "ts": 679622084284.487,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622084304.822, "dur": 16.211,
    "args": {
      "External id": 531004,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622084312.557, "dur": 7.832,
    "args": {
      "External id": 531005,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622084412.632, "dur": 109.586,
    "args": {
      "External id": 531006,"Record function id": 0, "Ev Idx": 5165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622084524.647, "dur": 1560.244,
    "args": {
      "External id": 531007,"Record function id": 0, "Ev Idx": 5166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622084577.577, "dur": 1486.122,
    "args": {
      "External id": 531008,"Sequence number": 6319426, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5167
    }
  },
  {
    "ph": "s", "id": 202, "pid": 4183438, "tid": 4183438, "ts": 679622084577.577,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622084722.223, "dur": 61.490,
    "args": {
      "External id": 531009,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622084805.273, "dur": 113.364,
    "args": {
      "External id": 531010,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622084933.142, "dur": 48.843,
    "args": {
      "External id": 531011,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622084995.934, "dur": 39.269,
    "args": {
      "External id": 531012,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622085073.072, "dur": 69.262,
    "args": {
      "External id": 531013,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622085173.695, "dur": 24.594,
    "args": {
      "External id": 531014,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622085225.988, "dur": 168.103,
    "args": {
      "External id": 531015,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622085289.213, "dur": 15.643,
    "args": {
      "External id": 531016,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622085295.550, "dur": 8.313,
    "args": {
      "External id": 531017,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622085309.611, "dur": 7.506,
    "args": {
      "External id": 531018,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622085318.982, "dur": 1.414,
    "args": {
      "External id": 531019,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622085323.250, "dur": 4.643,
    "args": {
      "External id": 531020,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622085410.108, "dur": 74.782,
    "args": {
      "External id": 531021,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622085526.837, "dur": 37.728,
    "args": {
      "External id": 531022,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622085576.672, "dur": 64.311,
    "args": {
      "External id": 531023,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622085651.250, "dur": 115.983,
    "args": {
      "External id": 531024,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622085807.422, "dur": 31.652,
    "args": {
      "External id": 531025,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622085849.365, "dur": 48.903,
    "args": {
      "External id": 531026,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622085926.585, "dur": 29.670,
    "args": {
      "External id": 531027,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5186
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.4)", "pid": 4183438, "tid": 4183438,
    "ts": 679622086204.219, "dur": 108.347,
    "args": {
      "External id": 531028,"Record function id": 0, "Ev Idx": 5187
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622086409.335, "dur": 68.581,
    "args": {
      "External id": 531029,"Record function id": 0, "Ev Idx": 5188
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.5)", "pid": 4183438, "tid": 4183438,
    "ts": 679622086490.641, "dur": 8452.231,
    "args": {
      "External id": 531030,"Record function id": 0, "Ev Idx": 5189
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.5)", "pid": 4183438, "tid": 4183438,
    "ts": 679622086505.065, "dur": 1368.430,
    "args": {
      "External id": 531031,"Record function id": 0, "Ev Idx": 5190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622086610.747, "dur": 12.355,
    "args": {
      "External id": 531032,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622086639.832, "dur": 113.484,
    "args": {
      "External id": 531033,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086648.733, "dur": 2.976,
    "args": {
      "External id": 531034,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086718.545, "dur": 1.006,
    "args": {
      "External id": 531035,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086723.153, "dur": 0.977,
    "args": {
      "External id": 531036,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086725.849, "dur": 0.367,
    "args": {
      "External id": 531037,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086729.739, "dur": 0.517,
    "args": {
      "External id": 531038,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086733.451, "dur": 0.532,
    "args": {
      "External id": 531039,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086735.037, "dur": 4.596,
    "args": {
      "External id": 531040,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086741.222, "dur": 0.304,
    "args": {
      "External id": 531041,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086745.169, "dur": 0.552,
    "args": {
      "External id": 531042,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622086769.193, "dur": 60.468,
    "args": {
      "External id": 531043,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622086878.486, "dur": 164.130,
    "args": {
      "External id": 531044,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622086897.239, "dur": 7.760,
    "args": {
      "External id": 531045,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622086913.706, "dur": 13.607,
    "args": {
      "External id": 531046,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622086919.562, "dur": 7.195,
    "args": {
      "External id": 531047,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086923.544, "dur": 1.102,
    "args": {
      "External id": 531048,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622086937.081, "dur": 37.942,
    "args": {
      "External id": 531049,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086941.090, "dur": 3.250,
    "args": {
      "External id": 531050,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086945.644, "dur": 0.579,
    "args": {
      "External id": 531051,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086947.659, "dur": 0.518,
    "args": {
      "External id": 531052,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086953.360, "dur": 1.764,
    "args": {
      "External id": 531053,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086956.241, "dur": 0.531,
    "args": {
      "External id": 531054,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086958.162, "dur": 0.448,
    "args": {
      "External id": 531055,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086962.950, "dur": 0.534,
    "args": {
      "External id": 531056,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086965.246, "dur": 0.440,
    "args": {
      "External id": 531057,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622086967.404, "dur": 2.726,
    "args": {
      "External id": 531058,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622086995.986, "dur": 35.026,
    "args": {
      "External id": 531059,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622087148.586, "dur": 578.213,
    "args": {
      "External id": 531060,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622087188.238, "dur": 528.623,
    "args": {
      "External id": 531061,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5220, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622087201.956, "dur": 438.664,
    "args": {
      "External id": 531062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622087765.197, "dur": 4.991,
    "args": {
      "External id": 531063,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5222, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.5)", "pid": 4183438, "tid": 4183438,
    "ts": 679622087904.101, "dur": 6816.598,
    "args": {
      "External id": 531064,"Record function id": 0, "Ev Idx": 5223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088033.594, "dur": 8.632,
    "args": {
      "External id": 531065,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088046.256, "dur": 1.402,
    "args": {
      "External id": 531066,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088051.955, "dur": 2.265,
    "args": {
      "External id": 531067,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088057.160, "dur": 1.014,
    "args": {
      "External id": 531068,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088060.528, "dur": 0.960,
    "args": {
      "External id": 531069,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088063.990, "dur": 1.409,
    "args": {
      "External id": 531070,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088070.370, "dur": 1.299,
    "args": {
      "External id": 531071,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088074.436, "dur": 1.956,
    "args": {
      "External id": 531072,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088079.031, "dur": 1.312,
    "args": {
      "External id": 531073,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622088082.436, "dur": 0.941,
    "args": {
      "External id": 531074,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622088143.577, "dur": 6474.541,
    "args": {
      "External id": 531075,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622088176.283, "dur": 6429.107,
    "args": {
      "External id": 531076,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622088197.820, "dur": 20.406,
    "args": {
      "External id": 531077,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622088221.653, "dur": 6341.702,
    "args": {
      "External id": 531078,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622088225.308, "dur": 6337.077,
    "args": {
      "External id": 531079,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622088233.138, "dur": 9.973,
    "args": {
      "External id": 531080,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622088245.558, "dur": 6308.135,
    "args": {
      "External id": 531081,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622094880.850, "dur": 34.391,
    "args": {
      "External id": 531082,"Sequence number": 6319427, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5241
    }
  },
  {
    "ph": "s", "id": 201, "pid": 4183438, "tid": 4183438, "ts": 679622094880.850,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622094898.558, "dur": 11.675,
    "args": {
      "External id": 531083,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622094903.387, "dur": 6.486,
    "args": {
      "External id": 531084,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622094983.256, "dur": 89.061,
    "args": {
      "External id": 531085,"Record function id": 0, "Ev Idx": 5244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622095074.673, "dur": 1172.776,
    "args": {
      "External id": 531086,"Record function id": 0, "Ev Idx": 5245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622095142.516, "dur": 1090.792,
    "args": {
      "External id": 531087,"Sequence number": 6319428, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5246
    }
  },
  {
    "ph": "s", "id": 200, "pid": 4183438, "tid": 4183438, "ts": 679622095142.516,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622095215.866, "dur": 47.242,
    "args": {
      "External id": 531088,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622095277.044, "dur": 93.268,
    "args": {
      "External id": 531089,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622095379.928, "dur": 34.910,
    "args": {
      "External id": 531090,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622095423.723, "dur": 29.420,
    "args": {
      "External id": 531091,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622095476.879, "dur": 25.423,
    "args": {
      "External id": 531092,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622095519.130, "dur": 14.980,
    "args": {
      "External id": 531093,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622095559.205, "dur": 189.502,
    "args": {
      "External id": 531094,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622095608.312, "dur": 12.501,
    "args": {
      "External id": 531095,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622095613.342, "dur": 6.448,
    "args": {
      "External id": 531096,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622095623.826, "dur": 3.962,
    "args": {
      "External id": 531097,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622095629.038, "dur": 1.134,
    "args": {
      "External id": 531098,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622095632.683, "dur": 3.662,
    "args": {
      "External id": 531099,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622095761.875, "dur": 52.296,
    "args": {
      "External id": 531100,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622095848.129, "dur": 30.438,
    "args": {
      "External id": 531101,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622095888.089, "dur": 40.464,
    "args": {
      "External id": 531102,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622095937.666, "dur": 48.304,
    "args": {
      "External id": 531103,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622096007.582, "dur": 31.764,
    "args": {
      "External id": 531104,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622096044.808, "dur": 51.486,
    "args": {
      "External id": 531105,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622096139.003, "dur": 23.791,
    "args": {
      "External id": 531106,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5265
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.5)", "pid": 4183438, "tid": 4183438,
    "ts": 679622096313.735, "dur": 75.072,
    "args": {
      "External id": 531107,"Record function id": 0, "Ev Idx": 5266
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622096462.229, "dur": 49.548,
    "args": {
      "External id": 531108,"Record function id": 0, "Ev Idx": 5267
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.6)", "pid": 4183438, "tid": 4183438,
    "ts": 679622096521.307, "dur": 8887.208,
    "args": {
      "External id": 531109,"Record function id": 0, "Ev Idx": 5268
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.6)", "pid": 4183438, "tid": 4183438,
    "ts": 679622096530.236, "dur": 932.666,
    "args": {
      "External id": 531110,"Record function id": 0, "Ev Idx": 5269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622096609.379, "dur": 8.555,
    "args": {
      "External id": 531111,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622096630.123, "dur": 81.781,
    "args": {
      "External id": 531112,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096635.937, "dur": 2.084,
    "args": {
      "External id": 531113,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096643.521, "dur": 0.479,
    "args": {
      "External id": 531114,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096645.933, "dur": 0.447,
    "args": {
      "External id": 531115,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096647.890, "dur": 0.427,
    "args": {
      "External id": 531116,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096692.008, "dur": 0.518,
    "args": {
      "External id": 531117,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096694.655, "dur": 0.376,
    "args": {
      "External id": 531118,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096696.620, "dur": 4.093,
    "args": {
      "External id": 531119,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096702.454, "dur": 0.405,
    "args": {
      "External id": 531120,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096704.528, "dur": 0.340,
    "args": {
      "External id": 531121,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622096723.598, "dur": 43.907,
    "args": {
      "External id": 531122,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622096806.112, "dur": 118.738,
    "args": {
      "External id": 531123,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622096819.582, "dur": 5.320,
    "args": {
      "External id": 531124,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622096830.890, "dur": 10.948,
    "args": {
      "External id": 531125,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622096835.355, "dur": 6.102,
    "args": {
      "External id": 531126,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096839.670, "dur": 0.515,
    "args": {
      "External id": 531127,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622096849.773, "dur": 31.423,
    "args": {
      "External id": 531128,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096852.133, "dur": 2.974,
    "args": {
      "External id": 531129,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096856.696, "dur": 0.398,
    "args": {
      "External id": 531130,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096858.407, "dur": 0.425,
    "args": {
      "External id": 531131,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096862.615, "dur": 1.433,
    "args": {
      "External id": 531132,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096865.582, "dur": 0.226,
    "args": {
      "External id": 531133,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096867.226, "dur": 0.361,
    "args": {
      "External id": 531134,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096870.532, "dur": 0.391,
    "args": {
      "External id": 531135,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096872.241, "dur": 0.358,
    "args": {
      "External id": 531136,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622096874.330, "dur": 2.797,
    "args": {
      "External id": 531137,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622096895.453, "dur": 21.426,
    "args": {
      "External id": 531138,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622096980.236, "dur": 385.862,
    "args": {
      "External id": 531139,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622097012.516, "dur": 348.143,
    "args": {
      "External id": 531140,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5299, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622097023.232, "dur": 331.452,
    "args": {
      "External id": 531141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622097389.378, "dur": 2.605,
    "args": {
      "External id": 531142,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5301, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.6)", "pid": 4183438, "tid": 4183438,
    "ts": 679622097484.920, "dur": 7727.253,
    "args": {
      "External id": 531143,"Record function id": 0, "Ev Idx": 5302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097576.364, "dur": 6.636,
    "args": {
      "External id": 531144,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097587.330, "dur": 0.881,
    "args": {
      "External id": 531145,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097590.109, "dur": 2.137,
    "args": {
      "External id": 531146,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097594.141, "dur": 0.716,
    "args": {
      "External id": 531147,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097596.569, "dur": 0.842,
    "args": {
      "External id": 531148,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097598.949, "dur": 0.615,
    "args": {
      "External id": 531149,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097602.670, "dur": 0.812,
    "args": {
      "External id": 531150,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097605.246, "dur": 1.878,
    "args": {
      "External id": 531151,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097608.537, "dur": 0.678,
    "args": {
      "External id": 531152,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622097610.784, "dur": 0.588,
    "args": {
      "External id": 531153,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622097630.968, "dur": 7530.280,
    "args": {
      "External id": 531154,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622097647.086, "dur": 7505.444,
    "args": {
      "External id": 531155,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622097704.307, "dur": 16.020,
    "args": {
      "External id": 531156,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622097723.384, "dur": 7390.910,
    "args": {
      "External id": 531157,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622097725.923, "dur": 7387.674,
    "args": {
      "External id": 531158,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622097732.678, "dur": 7.596,
    "args": {
      "External id": 531159,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622097742.247, "dur": 7367.753,
    "args": {
      "External id": 531160,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622105351.882, "dur": 31.363,
    "args": {
      "External id": 531161,"Sequence number": 6319429, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5320
    }
  },
  {
    "ph": "s", "id": 199, "pid": 4183438, "tid": 4183438, "ts": 679622105351.882,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622105368.870, "dur": 9.662,
    "args": {
      "External id": 531162,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622105372.585, "dur": 5.667,
    "args": {
      "External id": 531163,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622105447.562, "dur": 90.452,
    "args": {
      "External id": 531164,"Record function id": 0, "Ev Idx": 5323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622105539.870, "dur": 1073.884,
    "args": {
      "External id": 531165,"Record function id": 0, "Ev Idx": 5324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622105579.878, "dur": 1018.793,
    "args": {
      "External id": 531166,"Sequence number": 6319430, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5325
    }
  },
  {
    "ph": "s", "id": 198, "pid": 4183438, "tid": 4183438, "ts": 679622105579.878,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622105646.003, "dur": 81.472,
    "args": {
      "External id": 531167,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622105744.383, "dur": 95.177,
    "args": {
      "External id": 531168,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622105849.283, "dur": 35.885,
    "args": {
      "External id": 531169,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622105894.356, "dur": 29.349,
    "args": {
      "External id": 531170,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622105950.047, "dur": 27.314,
    "args": {
      "External id": 531171,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622105994.590, "dur": 13.357,
    "args": {
      "External id": 531172,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622106027.058, "dur": 153.909,
    "args": {
      "External id": 531173,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622106077.982, "dur": 11.210,
    "args": {
      "External id": 531174,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622106083.429, "dur": 5.052,
    "args": {
      "External id": 531175,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622106091.947, "dur": 23.687,
    "args": {
      "External id": 531176,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622106118.623, "dur": 1.278,
    "args": {
      "External id": 531177,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622106122.438, "dur": 3.668,
    "args": {
      "External id": 531178,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622106193.327, "dur": 49.991,
    "args": {
      "External id": 531179,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622106271.372, "dur": 30.118,
    "args": {
      "External id": 531180,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622106310.428, "dur": 37.884,
    "args": {
      "External id": 531181,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622106357.233, "dur": 32.049,
    "args": {
      "External id": 531182,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622106411.097, "dur": 25.296,
    "args": {
      "External id": 531183,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622106441.720, "dur": 45.198,
    "args": {
      "External id": 531184,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622106506.310, "dur": 20.853,
    "args": {
      "External id": 531185,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5344
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.6)", "pid": 4183438, "tid": 4183438,
    "ts": 679622106720.254, "dur": 79.360,
    "args": {
      "External id": 531186,"Record function id": 0, "Ev Idx": 5345
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622106875.938, "dur": 48.010,
    "args": {
      "External id": 531187,"Record function id": 0, "Ev Idx": 5346
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.7)", "pid": 4183438, "tid": 4183438,
    "ts": 679622106933.043, "dur": 9501.827,
    "args": {
      "External id": 531188,"Record function id": 0, "Ev Idx": 5347
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.7)", "pid": 4183438, "tid": 4183438,
    "ts": 679622106942.902, "dur": 945.484,
    "args": {
      "External id": 531189,"Record function id": 0, "Ev Idx": 5348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622107021.844, "dur": 8.656,
    "args": {
      "External id": 531190,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622107044.588, "dur": 39.183,
    "args": {
      "External id": 531191,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107050.030, "dur": 2.450,
    "args": {
      "External id": 531192,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107057.237, "dur": 0.401,
    "args": {
      "External id": 531193,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107059.414, "dur": 0.455,
    "args": {
      "External id": 531194,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107061.132, "dur": 0.226,
    "args": {
      "External id": 531195,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107065.379, "dur": 0.507,
    "args": {
      "External id": 531196,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107067.838, "dur": 0.201,
    "args": {
      "External id": 531197,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107069.535, "dur": 3.867,
    "args": {
      "External id": 531198,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107075.095, "dur": 0.400,
    "args": {
      "External id": 531199,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107077.023, "dur": 0.430,
    "args": {
      "External id": 531200,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622107094.957, "dur": 58.954,
    "args": {
      "External id": 531201,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622107192.196, "dur": 119.814,
    "args": {
      "External id": 531202,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622107203.536, "dur": 4.843,
    "args": {
      "External id": 531203,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622107213.740, "dur": 11.505,
    "args": {
      "External id": 531204,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622107218.682, "dur": 6.141,
    "args": {
      "External id": 531205,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107222.666, "dur": 0.701,
    "args": {
      "External id": 531206,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622107232.129, "dur": 33.492,
    "args": {
      "External id": 531207,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107234.697, "dur": 2.706,
    "args": {
      "External id": 531208,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107238.948, "dur": 0.545,
    "args": {
      "External id": 531209,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107241.284, "dur": 0.531,
    "args": {
      "External id": 531210,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107246.322, "dur": 1.542,
    "args": {
      "External id": 531211,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107249.330, "dur": 0.484,
    "args": {
      "External id": 531212,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107251.414, "dur": 0.289,
    "args": {
      "External id": 531213,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107255.305, "dur": 0.461,
    "args": {
      "External id": 531214,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107257.297, "dur": 0.164,
    "args": {
      "External id": 531215,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622107258.681, "dur": 2.032,
    "args": {
      "External id": 531216,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622107281.714, "dur": 22.290,
    "args": {
      "External id": 531217,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622107365.574, "dur": 421.335,
    "args": {
      "External id": 531218,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622107398.618, "dur": 382.700,
    "args": {
      "External id": 531219,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5378, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622107409.450, "dur": 365.114,
    "args": {
      "External id": 531220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622107811.526, "dur": 2.703,
    "args": {
      "External id": 531221,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5380, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.7)", "pid": 4183438, "tid": 4183438,
    "ts": 679622107909.125, "dur": 8300.031,
    "args": {
      "External id": 531222,"Record function id": 0, "Ev Idx": 5381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108011.716, "dur": 7.277,
    "args": {
      "External id": 531223,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108022.810, "dur": 1.198,
    "args": {
      "External id": 531224,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108025.831, "dur": 2.660,
    "args": {
      "External id": 531225,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108031.189, "dur": 0.770,
    "args": {
      "External id": 531226,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108034.163, "dur": 1.007,
    "args": {
      "External id": 531227,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108036.909, "dur": 0.898,
    "args": {
      "External id": 531228,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108039.383, "dur": 0.764,
    "args": {
      "External id": 531229,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108043.385, "dur": 2.054,
    "args": {
      "External id": 531230,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108046.826, "dur": 0.685,
    "args": {
      "External id": 531231,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622108049.065, "dur": 0.760,
    "args": {
      "External id": 531232,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622108069.358, "dur": 8067.868,
    "args": {
      "External id": 531233,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622108085.184, "dur": 8038.135,
    "args": {
      "External id": 531234,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622108116.385, "dur": 16.053,
    "args": {
      "External id": 531235,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622108137.127, "dur": 7919.935,
    "args": {
      "External id": 531236,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622108139.998, "dur": 7915.835,
    "args": {
      "External id": 531237,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622108146.931, "dur": 7.541,
    "args": {
      "External id": 531238,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622108156.423, "dur": 7893.032,
    "args": {
      "External id": 531239,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622116371.012, "dur": 36.492,
    "args": {
      "External id": 531240,"Sequence number": 6319431, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5399
    }
  },
  {
    "ph": "s", "id": 197, "pid": 4183438, "tid": 4183438, "ts": 679622116371.012,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622116390.151, "dur": 12.492,
    "args": {
      "External id": 531241,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622116395.624, "dur": 6.775,
    "args": {
      "External id": 531242,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622116473.798, "dur": 88.574,
    "args": {
      "External id": 531243,"Record function id": 0, "Ev Idx": 5402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622116565.094, "dur": 1316.749,
    "args": {
      "External id": 531244,"Record function id": 0, "Ev Idx": 5403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622116609.706, "dur": 1252.248,
    "args": {
      "External id": 531245,"Sequence number": 6319432, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5404
    }
  },
  {
    "ph": "s", "id": 196, "pid": 4183438, "tid": 4183438, "ts": 679622116609.706,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622116725.730, "dur": 49.327,
    "args": {
      "External id": 531246,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622116791.243, "dur": 97.262,
    "args": {
      "External id": 531247,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622116898.853, "dur": 36.372,
    "args": {
      "External id": 531248,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622116944.891, "dur": 29.704,
    "args": {
      "External id": 531249,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622116997.606, "dur": 26.324,
    "args": {
      "External id": 531250,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622117044.056, "dur": 17.154,
    "args": {
      "External id": 531251,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622117081.418, "dur": 177.566,
    "args": {
      "External id": 531252,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622117163.574, "dur": 13.930,
    "args": {
      "External id": 531253,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622117169.873, "dur": 6.406,
    "args": {
      "External id": 531254,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622117180.331, "dur": 6.200,
    "args": {
      "External id": 531255,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622117196.271, "dur": 1.422,
    "args": {
      "External id": 531256,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622117200.108, "dur": 4.030,
    "args": {
      "External id": 531257,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622117271.002, "dur": 53.947,
    "args": {
      "External id": 531258,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622117358.368, "dur": 31.865,
    "args": {
      "External id": 531259,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622117399.969, "dur": 40.582,
    "args": {
      "External id": 531260,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622117449.054, "dur": 52.685,
    "args": {
      "External id": 531261,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622117531.437, "dur": 46.675,
    "args": {
      "External id": 531262,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622117588.219, "dur": 56.703,
    "args": {
      "External id": 531263,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622117731.072, "dur": 31.444,
    "args": {
      "External id": 531264,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5423
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.7)", "pid": 4183438, "tid": 4183438,
    "ts": 679622117969.092, "dur": 102.347,
    "args": {
      "External id": 531265,"Record function id": 0, "Ev Idx": 5424
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622118181.242, "dur": 51.174,
    "args": {
      "External id": 531266,"Record function id": 0, "Ev Idx": 5425
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.8)", "pid": 4183438, "tid": 4183438,
    "ts": 679622118242.269, "dur": 8916.528,
    "args": {
      "External id": 531267,"Record function id": 0, "Ev Idx": 5426
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.8)", "pid": 4183438, "tid": 4183438,
    "ts": 679622118251.093, "dur": 934.925,
    "args": {
      "External id": 531268,"Record function id": 0, "Ev Idx": 5427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622118334.959, "dur": 11.556,
    "args": {
      "External id": 531269,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622118360.732, "dur": 41.284,
    "args": {
      "External id": 531270,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118366.367, "dur": 2.613,
    "args": {
      "External id": 531271,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118374.192, "dur": 0.274,
    "args": {
      "External id": 531272,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118376.348, "dur": 0.424,
    "args": {
      "External id": 531273,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118378.450, "dur": 0.862,
    "args": {
      "External id": 531274,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118382.702, "dur": 0.546,
    "args": {
      "External id": 531275,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118384.740, "dur": 0.430,
    "args": {
      "External id": 531276,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118387.356, "dur": 3.722,
    "args": {
      "External id": 531277,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118392.966, "dur": 0.402,
    "args": {
      "External id": 531278,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118394.991, "dur": 0.481,
    "args": {
      "External id": 531279,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622118412.779, "dur": 44.333,
    "args": {
      "External id": 531280,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622118491.720, "dur": 117.243,
    "args": {
      "External id": 531281,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622118502.676, "dur": 4.464,
    "args": {
      "External id": 531282,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622118512.837, "dur": 11.163,
    "args": {
      "External id": 531283,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622118517.389, "dur": 6.204,
    "args": {
      "External id": 531284,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118521.841, "dur": 0.375,
    "args": {
      "External id": 531285,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622118530.839, "dur": 34.919,
    "args": {
      "External id": 531286,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118533.302, "dur": 3.085,
    "args": {
      "External id": 531287,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118538.389, "dur": 0.461,
    "args": {
      "External id": 531288,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118540.532, "dur": 0.352,
    "args": {
      "External id": 531289,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118544.632, "dur": 1.549,
    "args": {
      "External id": 531290,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118547.849, "dur": 0.447,
    "args": {
      "External id": 531291,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118549.745, "dur": 0.506,
    "args": {
      "External id": 531292,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118554.285, "dur": 0.451,
    "args": {
      "External id": 531293,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118556.339, "dur": 0.528,
    "args": {
      "External id": 531294,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622118558.509, "dur": 2.992,
    "args": {
      "External id": 531295,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622118578.173, "dur": 22.852,
    "args": {
      "External id": 531296,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622118700.848, "dur": 363.098,
    "args": {
      "External id": 531297,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622118732.829, "dur": 325.792,
    "args": {
      "External id": 531298,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5457, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622118744.711, "dur": 308.293,
    "args": {
      "External id": 531299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622119084.844, "dur": 2.459,
    "args": {
      "External id": 531300,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5459, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.8)", "pid": 4183438, "tid": 4183438,
    "ts": 679622119209.632, "dur": 7659.467,
    "args": {
      "External id": 531301,"Record function id": 0, "Ev Idx": 5460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119307.805, "dur": 6.939,
    "args": {
      "External id": 531302,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119318.493, "dur": 1.071,
    "args": {
      "External id": 531303,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119321.287, "dur": 2.544,
    "args": {
      "External id": 531304,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119325.836, "dur": 1.043,
    "args": {
      "External id": 531305,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119328.523, "dur": 0.944,
    "args": {
      "External id": 531306,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119331.449, "dur": 1.043,
    "args": {
      "External id": 531307,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119334.157, "dur": 1.009,
    "args": {
      "External id": 531308,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119337.269, "dur": 2.202,
    "args": {
      "External id": 531309,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119341.436, "dur": 0.889,
    "args": {
      "External id": 531310,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622119343.933, "dur": 0.696,
    "args": {
      "External id": 531311,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622119375.687, "dur": 7425.498,
    "args": {
      "External id": 531312,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622119396.323, "dur": 7394.541,
    "args": {
      "External id": 531313,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622119411.027, "dur": 15.748,
    "args": {
      "External id": 531314,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622119429.435, "dur": 7307.335,
    "args": {
      "External id": 531315,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622119432.323, "dur": 7303.656,
    "args": {
      "External id": 531316,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622119438.730, "dur": 7.374,
    "args": {
      "External id": 531317,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622119448.024, "dur": 7283.995,
    "args": {
      "External id": 531318,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622127050.945, "dur": 40.525,
    "args": {
      "External id": 531319,"Sequence number": 6319433, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5478
    }
  },
  {
    "ph": "s", "id": 195, "pid": 4183438, "tid": 4183438, "ts": 679622127050.945,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622127072.484, "dur": 13.203,
    "args": {
      "External id": 531320,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622127078.237, "dur": 7.019,
    "args": {
      "External id": 531321,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622127209.204, "dur": 111.288,
    "args": {
      "External id": 531322,"Record function id": 0, "Ev Idx": 5481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622127323.394, "dur": 1457.370,
    "args": {
      "External id": 531323,"Record function id": 0, "Ev Idx": 5482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622127377.339, "dur": 1386.785,
    "args": {
      "External id": 531324,"Sequence number": 6319434, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5483
    }
  },
  {
    "ph": "s", "id": 194, "pid": 4183438, "tid": 4183438, "ts": 679622127377.339,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622127471.413, "dur": 65.132,
    "args": {
      "External id": 531325,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622127555.271, "dur": 179.883,
    "args": {
      "External id": 531326,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622127756.647, "dur": 57.375,
    "args": {
      "External id": 531327,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622127830.005, "dur": 41.784,
    "args": {
      "External id": 531328,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622127911.609, "dur": 36.096,
    "args": {
      "External id": 531329,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622127975.012, "dur": 28.216,
    "args": {
      "External id": 531330,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622128030.015, "dur": 201.762,
    "args": {
      "External id": 531331,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622128094.037, "dur": 40.058,
    "args": {
      "External id": 531332,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622128123.031, "dur": 9.684,
    "args": {
      "External id": 531333,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622128138.315, "dur": 6.742,
    "args": {
      "External id": 531334,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622128147.491, "dur": 1.990,
    "args": {
      "External id": 531335,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622128152.615, "dur": 6.071,
    "args": {
      "External id": 531336,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622128247.877, "dur": 76.190,
    "args": {
      "External id": 531337,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622128365.506, "dur": 40.355,
    "args": {
      "External id": 531338,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622128418.462, "dur": 49.212,
    "args": {
      "External id": 531339,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622128477.205, "dur": 33.920,
    "args": {
      "External id": 531340,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622128534.378, "dur": 26.849,
    "args": {
      "External id": 531341,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622128566.529, "dur": 32.782,
    "args": {
      "External id": 531342,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622128619.458, "dur": 18.323,
    "args": {
      "External id": 531343,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5502
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.8)", "pid": 4183438, "tid": 4183438,
    "ts": 679622128851.098, "dur": 79.098,
    "args": {
      "External id": 531344,"Record function id": 0, "Ev Idx": 5503
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622129004.301, "dur": 49.526,
    "args": {
      "External id": 531345,"Record function id": 0, "Ev Idx": 5504
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.9)", "pid": 4183438, "tid": 4183438,
    "ts": 679622129065.043, "dur": 8346.738,
    "args": {
      "External id": 531346,"Record function id": 0, "Ev Idx": 5505
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.9)", "pid": 4183438, "tid": 4183438,
    "ts": 679622129079.922, "dur": 944.749,
    "args": {
      "External id": 531347,"Record function id": 0, "Ev Idx": 5506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622129188.070, "dur": 11.173,
    "args": {
      "External id": 531348,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622129213.147, "dur": 35.884,
    "args": {
      "External id": 531349,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129218.211, "dur": 2.107,
    "args": {
      "External id": 531350,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129224.990, "dur": 0.245,
    "args": {
      "External id": 531351,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129226.293, "dur": 0.463,
    "args": {
      "External id": 531352,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129228.280, "dur": 0.658,
    "args": {
      "External id": 531353,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129232.397, "dur": 0.565,
    "args": {
      "External id": 531354,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129233.779, "dur": 0.510,
    "args": {
      "External id": 531355,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129235.192, "dur": 4.124,
    "args": {
      "External id": 531356,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129240.628, "dur": 0.217,
    "args": {
      "External id": 531357,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129241.984, "dur": 0.325,
    "args": {
      "External id": 531358,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622129260.149, "dur": 44.973,
    "args": {
      "External id": 531359,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622129339.828, "dur": 127.929,
    "args": {
      "External id": 531360,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622129350.424, "dur": 4.911,
    "args": {
      "External id": 531361,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622129360.769, "dur": 10.299,
    "args": {
      "External id": 531362,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622129365.415, "dur": 5.255,
    "args": {
      "External id": 531363,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129368.701, "dur": 0.649,
    "args": {
      "External id": 531364,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622129377.468, "dur": 36.652,
    "args": {
      "External id": 531365,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129379.409, "dur": 2.538,
    "args": {
      "External id": 531366,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129390.624, "dur": 0.408,
    "args": {
      "External id": 531367,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129392.131, "dur": 0.415,
    "args": {
      "External id": 531368,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129396.282, "dur": 2.256,
    "args": {
      "External id": 531369,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129399.500, "dur": 0.415,
    "args": {
      "External id": 531370,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129400.863, "dur": 2.232,
    "args": {
      "External id": 531371,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129404.010, "dur": 0.385,
    "args": {
      "External id": 531372,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129405.513, "dur": 0.187,
    "args": {
      "External id": 531373,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622129409.137, "dur": 0.190,
    "args": {
      "External id": 531374,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622129434.958, "dur": 24.760,
    "args": {
      "External id": 531375,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622129521.023, "dur": 404.357,
    "args": {
      "External id": 531376,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622129548.182, "dur": 371.499,
    "args": {
      "External id": 531377,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5536, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622129560.370, "dur": 353.107,
    "args": {
      "External id": 531378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622129947.628, "dur": 2.319,
    "args": {
      "External id": 531379,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5538, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.9)", "pid": 4183438, "tid": 4183438,
    "ts": 679622130047.257, "dur": 7176.597,
    "args": {
      "External id": 531380,"Record function id": 0, "Ev Idx": 5539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130160.476, "dur": 7.271,
    "args": {
      "External id": 531381,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130171.786, "dur": 1.352,
    "args": {
      "External id": 531382,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130174.810, "dur": 2.805,
    "args": {
      "External id": 531383,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130179.341, "dur": 1.059,
    "args": {
      "External id": 531384,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130182.095, "dur": 0.918,
    "args": {
      "External id": 531385,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130184.790, "dur": 1.026,
    "args": {
      "External id": 531386,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130187.560, "dur": 0.966,
    "args": {
      "External id": 531387,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130190.182, "dur": 2.488,
    "args": {
      "External id": 531388,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130194.186, "dur": 0.631,
    "args": {
      "External id": 531389,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622130196.034, "dur": 0.886,
    "args": {
      "External id": 531390,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622130216.020, "dur": 6968.882,
    "args": {
      "External id": 531391,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622130232.732, "dur": 6945.432,
    "args": {
      "External id": 531392,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622130249.004, "dur": 15.980,
    "args": {
      "External id": 531393,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622130267.780, "dur": 6877.675,
    "args": {
      "External id": 531394,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622130270.610, "dur": 6874.288,
    "args": {
      "External id": 531395,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622130277.759, "dur": 7.553,
    "args": {
      "External id": 531396,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622130287.184, "dur": 6854.569,
    "args": {
      "External id": 531397,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622137357.176, "dur": 29.423,
    "args": {
      "External id": 531398,"Sequence number": 6319435, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5557
    }
  },
  {
    "ph": "s", "id": 193, "pid": 4183438, "tid": 4183438, "ts": 679622137357.176,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622137373.548, "dur": 8.017,
    "args": {
      "External id": 531399,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622137376.919, "dur": 4.387,
    "args": {
      "External id": 531400,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622137451.945, "dur": 87.782,
    "args": {
      "External id": 531401,"Record function id": 0, "Ev Idx": 5560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622137541.455, "dur": 1069.309,
    "args": {
      "External id": 531402,"Record function id": 0, "Ev Idx": 5561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622137580.526, "dur": 1016.201,
    "args": {
      "External id": 531403,"Sequence number": 6319436, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5562
    }
  },
  {
    "ph": "s", "id": 192, "pid": 4183438, "tid": 4183438, "ts": 679622137580.526,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622137644.489, "dur": 79.441,
    "args": {
      "External id": 531404,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622137740.569, "dur": 94.775,
    "args": {
      "External id": 531405,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622137846.393, "dur": 35.166,
    "args": {
      "External id": 531406,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622137890.271, "dur": 30.114,
    "args": {
      "External id": 531407,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622137944.076, "dur": 26.034,
    "args": {
      "External id": 531408,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622137988.230, "dur": 17.620,
    "args": {
      "External id": 531409,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622138025.150, "dur": 155.704,
    "args": {
      "External id": 531410,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622138076.105, "dur": 11.866,
    "args": {
      "External id": 531411,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622138081.598, "dur": 5.403,
    "args": {
      "External id": 531412,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622138090.795, "dur": 5.486,
    "args": {
      "External id": 531413,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622138115.561, "dur": 2.333,
    "args": {
      "External id": 531414,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622138121.127, "dur": 4.369,
    "args": {
      "External id": 531415,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622138192.402, "dur": 49.383,
    "args": {
      "External id": 531416,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622138272.474, "dur": 27.400,
    "args": {
      "External id": 531417,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622138316.152, "dur": 39.389,
    "args": {
      "External id": 531418,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622138364.503, "dur": 31.612,
    "args": {
      "External id": 531419,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622138417.040, "dur": 26.005,
    "args": {
      "External id": 531420,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622138448.644, "dur": 41.691,
    "args": {
      "External id": 531421,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622138510.998, "dur": 17.966,
    "args": {
      "External id": 531422,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5581
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.9)", "pid": 4183438, "tid": 4183438,
    "ts": 679622138721.426, "dur": 81.358,
    "args": {
      "External id": 531423,"Record function id": 0, "Ev Idx": 5582
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622138879.381, "dur": 47.043,
    "args": {
      "External id": 531424,"Record function id": 0, "Ev Idx": 5583
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.10)", "pid": 4183438, "tid": 4183438,
    "ts": 679622138935.909, "dur": 8455.000,
    "args": {
      "External id": 531425,"Record function id": 0, "Ev Idx": 5584
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.10)", "pid": 4183438, "tid": 4183438,
    "ts": 679622138945.795, "dur": 1022.447,
    "args": {
      "External id": 531426,"Record function id": 0, "Ev Idx": 5585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622139025.779, "dur": 9.840,
    "args": {
      "External id": 531427,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622139048.931, "dur": 37.986,
    "args": {
      "External id": 531428,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139054.678, "dur": 1.986,
    "args": {
      "External id": 531429,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139061.431, "dur": 0.222,
    "args": {
      "External id": 531430,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139063.335, "dur": 0.596,
    "args": {
      "External id": 531431,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139065.563, "dur": 0.435,
    "args": {
      "External id": 531432,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139069.500, "dur": 0.551,
    "args": {
      "External id": 531433,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139071.364, "dur": 0.716,
    "args": {
      "External id": 531434,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139073.473, "dur": 2.833,
    "args": {
      "External id": 531435,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139077.785, "dur": 0.386,
    "args": {
      "External id": 531436,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139079.587, "dur": 0.219,
    "args": {
      "External id": 531437,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622139114.122, "dur": 41.029,
    "args": {
      "External id": 531438,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622139192.621, "dur": 115.035,
    "args": {
      "External id": 531439,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622139204.131, "dur": 4.990,
    "args": {
      "External id": 531440,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622139214.233, "dur": 11.418,
    "args": {
      "External id": 531441,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622139218.799, "dur": 6.432,
    "args": {
      "External id": 531442,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139222.695, "dur": 0.922,
    "args": {
      "External id": 531443,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622139233.052, "dur": 32.006,
    "args": {
      "External id": 531444,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139235.991, "dur": 2.405,
    "args": {
      "External id": 531445,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139239.867, "dur": 0.421,
    "args": {
      "External id": 531446,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139242.031, "dur": 0.247,
    "args": {
      "External id": 531447,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139246.053, "dur": 1.424,
    "args": {
      "External id": 531448,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139249.218, "dur": 0.217,
    "args": {
      "External id": 531449,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139251.026, "dur": 0.449,
    "args": {
      "External id": 531450,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139254.648, "dur": 0.163,
    "args": {
      "External id": 531451,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139255.951, "dur": 0.169,
    "args": {
      "External id": 531452,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622139257.679, "dur": 2.361,
    "args": {
      "External id": 531453,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622139276.880, "dur": 22.901,
    "args": {
      "External id": 531454,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622139361.420, "dur": 502.022,
    "args": {
      "External id": 531455,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622139389.363, "dur": 468.077,
    "args": {
      "External id": 531456,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5615, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622139432.642, "dur": 418.173,
    "args": {
      "External id": 531457,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622139889.511, "dur": 2.834,
    "args": {
      "External id": 531458,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5617, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.10)", "pid": 4183438, "tid": 4183438,
    "ts": 679622139989.782, "dur": 7208.064,
    "args": {
      "External id": 531459,"Record function id": 0, "Ev Idx": 5618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140091.695, "dur": 23.447,
    "args": {
      "External id": 531460,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140120.826, "dur": 1.602,
    "args": {
      "External id": 531461,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140124.185, "dur": 2.990,
    "args": {
      "External id": 531462,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140128.714, "dur": 1.389,
    "args": {
      "External id": 531463,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140131.593, "dur": 1.117,
    "args": {
      "External id": 531464,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140134.592, "dur": 0.900,
    "args": {
      "External id": 531465,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140137.664, "dur": 0.896,
    "args": {
      "External id": 531466,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140140.128, "dur": 1.722,
    "args": {
      "External id": 531467,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140144.032, "dur": 0.746,
    "args": {
      "External id": 531468,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622140146.823, "dur": 0.582,
    "args": {
      "External id": 531469,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622140167.265, "dur": 6988.225,
    "args": {
      "External id": 531470,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622140184.817, "dur": 6963.124,
    "args": {
      "External id": 531471,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622140203.086, "dur": 15.439,
    "args": {
      "External id": 531472,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622140221.483, "dur": 6892.143,
    "args": {
      "External id": 531473,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622140224.296, "dur": 6888.712,
    "args": {
      "External id": 531474,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622140230.856, "dur": 7.370,
    "args": {
      "External id": 531475,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622140240.208, "dur": 6869.473,
    "args": {
      "External id": 531476,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622147332.627, "dur": 33.633,
    "args": {
      "External id": 531477,"Sequence number": 6319437, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5636
    }
  },
  {
    "ph": "s", "id": 191, "pid": 4183438, "tid": 4183438, "ts": 679622147332.627,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622147352.994, "dur": 8.667,
    "args": {
      "External id": 531478,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622147356.603, "dur": 4.752,
    "args": {
      "External id": 531479,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622147428.802, "dur": 90.151,
    "args": {
      "External id": 531480,"Record function id": 0, "Ev Idx": 5639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622147520.938, "dur": 1075.945,
    "args": {
      "External id": 531481,"Record function id": 0, "Ev Idx": 5640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622147561.261, "dur": 1022.130,
    "args": {
      "External id": 531482,"Sequence number": 6319438, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5641
    }
  },
  {
    "ph": "s", "id": 190, "pid": 4183438, "tid": 4183438, "ts": 679622147561.261,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622147627.751, "dur": 94.262,
    "args": {
      "External id": 531483,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622147739.150, "dur": 92.594,
    "args": {
      "External id": 531484,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622147842.583, "dur": 33.821,
    "args": {
      "External id": 531485,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622147885.682, "dur": 29.170,
    "args": {
      "External id": 531486,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622147939.872, "dur": 25.159,
    "args": {
      "External id": 531487,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622147981.029, "dur": 13.903,
    "args": {
      "External id": 531488,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622148013.652, "dur": 150.176,
    "args": {
      "External id": 531489,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622148064.762, "dur": 11.338,
    "args": {
      "External id": 531490,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622148069.766, "dur": 5.395,
    "args": {
      "External id": 531491,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622148078.912, "dur": 5.880,
    "args": {
      "External id": 531492,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622148086.299, "dur": 1.281,
    "args": {
      "External id": 531493,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622148090.431, "dur": 4.019,
    "args": {
      "External id": 531494,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622148176.648, "dur": 47.527,
    "args": {
      "External id": 531495,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622148254.733, "dur": 28.434,
    "args": {
      "External id": 531496,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622148292.077, "dur": 37.969,
    "args": {
      "External id": 531497,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622148338.971, "dur": 30.735,
    "args": {
      "External id": 531498,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622148391.010, "dur": 23.808,
    "args": {
      "External id": 531499,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622148419.877, "dur": 51.757,
    "args": {
      "External id": 531500,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622148490.431, "dur": 18.599,
    "args": {
      "External id": 531501,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5660
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.10)", "pid": 4183438, "tid": 4183438,
    "ts": 679622148701.621, "dur": 81.417,
    "args": {
      "External id": 531502,"Record function id": 0, "Ev Idx": 5661
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622148860.184, "dur": 49.823,
    "args": {
      "External id": 531503,"Record function id": 0, "Ev Idx": 5662
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.11)", "pid": 4183438, "tid": 4183438,
    "ts": 679622148919.566, "dur": 8956.005,
    "args": {
      "External id": 531504,"Record function id": 0, "Ev Idx": 5663
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.11)", "pid": 4183438, "tid": 4183438,
    "ts": 679622148928.541, "dur": 945.583,
    "args": {
      "External id": 531505,"Record function id": 0, "Ev Idx": 5664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622149009.128, "dur": 8.873,
    "args": {
      "External id": 531506,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622149031.781, "dur": 40.065,
    "args": {
      "External id": 531507,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149037.508, "dur": 2.449,
    "args": {
      "External id": 531508,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149044.975, "dur": 0.451,
    "args": {
      "External id": 531509,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149047.213, "dur": 0.570,
    "args": {
      "External id": 531510,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149049.137, "dur": 0.379,
    "args": {
      "External id": 531511,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149053.393, "dur": 0.409,
    "args": {
      "External id": 531512,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149055.393, "dur": 0.407,
    "args": {
      "External id": 531513,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149057.336, "dur": 3.895,
    "args": {
      "External id": 531514,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149062.688, "dur": 0.212,
    "args": {
      "External id": 531515,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149064.661, "dur": 0.446,
    "args": {
      "External id": 531516,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622149086.637, "dur": 56.673,
    "args": {
      "External id": 531517,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622149180.988, "dur": 123.006,
    "args": {
      "External id": 531518,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622149192.464, "dur": 7.521,
    "args": {
      "External id": 531519,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622149205.938, "dur": 12.497,
    "args": {
      "External id": 531520,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622149210.420, "dur": 7.588,
    "args": {
      "External id": 531521,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149215.250, "dur": 0.764,
    "args": {
      "External id": 531522,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622149226.595, "dur": 32.191,
    "args": {
      "External id": 531523,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149229.447, "dur": 0.616,
    "args": {
      "External id": 531524,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149232.231, "dur": 2.322,
    "args": {
      "External id": 531525,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149236.088, "dur": 0.383,
    "args": {
      "External id": 531526,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149238.174, "dur": 1.474,
    "args": {
      "External id": 531527,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149243.505, "dur": 0.224,
    "args": {
      "External id": 531528,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149245.084, "dur": 0.501,
    "args": {
      "External id": 531529,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149246.997, "dur": 0.459,
    "args": {
      "External id": 531530,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149250.952, "dur": 0.467,
    "args": {
      "External id": 531531,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622149253.663, "dur": 0.420,
    "args": {
      "External id": 531532,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622149272.255, "dur": 23.619,
    "args": {
      "External id": 531533,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622149358.156, "dur": 415.432,
    "args": {
      "External id": 531534,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622149386.076, "dur": 381.726,
    "args": {
      "External id": 531535,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5694, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622149396.352, "dur": 365.391,
    "args": {
      "External id": 531536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622149796.430, "dur": 2.562,
    "args": {
      "External id": 531537,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5696, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.11)", "pid": 4183438, "tid": 4183438,
    "ts": 679622149897.735, "dur": 7733.539,
    "args": {
      "External id": 531538,"Record function id": 0, "Ev Idx": 5697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150003.710, "dur": 7.130,
    "args": {
      "External id": 531539,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150014.192, "dur": 1.367,
    "args": {
      "External id": 531540,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150017.268, "dur": 2.299,
    "args": {
      "External id": 531541,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150021.207, "dur": 0.990,
    "args": {
      "External id": 531542,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150023.869, "dur": 0.765,
    "args": {
      "External id": 531543,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150026.272, "dur": 0.924,
    "args": {
      "External id": 531544,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150028.861, "dur": 0.816,
    "args": {
      "External id": 531545,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150031.568, "dur": 2.173,
    "args": {
      "External id": 531546,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150035.491, "dur": 0.832,
    "args": {
      "External id": 531547,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622150038.476, "dur": 0.641,
    "args": {
      "External id": 531548,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622150058.588, "dur": 7519.878,
    "args": {
      "External id": 531549,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622150075.042, "dur": 7493.099,
    "args": {
      "External id": 531550,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622150094.989, "dur": 31.748,
    "args": {
      "External id": 531551,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622150130.929, "dur": 7398.641,
    "args": {
      "External id": 531552,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622150133.559, "dur": 7395.267,
    "args": {
      "External id": 531553,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622150139.840, "dur": 6.458,
    "args": {
      "External id": 531554,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622150148.028, "dur": 7377.409,
    "args": {
      "External id": 531555,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622157808.135, "dur": 39.857,
    "args": {
      "External id": 531556,"Sequence number": 6319439, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5715
    }
  },
  {
    "ph": "s", "id": 189, "pid": 4183438, "tid": 4183438, "ts": 679622157808.135,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622157831.691, "dur": 10.791,
    "args": {
      "External id": 531557,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622157836.136, "dur": 5.986,
    "args": {
      "External id": 531558,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622157914.852, "dur": 92.810,
    "args": {
      "External id": 531559,"Record function id": 0, "Ev Idx": 5718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622158009.730, "dur": 1120.864,
    "args": {
      "External id": 531560,"Record function id": 0, "Ev Idx": 5719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622158052.080, "dur": 1062.242,
    "args": {
      "External id": 531561,"Sequence number": 6319440, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5720
    }
  },
  {
    "ph": "s", "id": 188, "pid": 4183438, "tid": 4183438, "ts": 679622158052.080,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622158135.284, "dur": 47.421,
    "args": {
      "External id": 531562,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622158198.447, "dur": 93.819,
    "args": {
      "External id": 531563,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622158301.578, "dur": 35.014,
    "args": {
      "External id": 531564,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622158345.627, "dur": 29.545,
    "args": {
      "External id": 531565,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622158399.731, "dur": 26.533,
    "args": {
      "External id": 531566,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622158443.444, "dur": 17.057,
    "args": {
      "External id": 531567,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622158479.069, "dur": 135.228,
    "args": {
      "External id": 531568,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622158531.787, "dur": 12.280,
    "args": {
      "External id": 531569,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622158537.517, "dur": 5.754,
    "args": {
      "External id": 531570,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622158547.292, "dur": 5.325,
    "args": {
      "External id": 531571,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622158554.144, "dur": 1.192,
    "args": {
      "External id": 531572,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622158558.110, "dur": 3.515,
    "args": {
      "External id": 531573,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622158625.509, "dur": 85.514,
    "args": {
      "External id": 531574,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622158746.099, "dur": 31.027,
    "args": {
      "External id": 531575,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622158786.822, "dur": 42.949,
    "args": {
      "External id": 531576,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622158839.182, "dur": 37.424,
    "args": {
      "External id": 531577,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622158898.963, "dur": 27.496,
    "args": {
      "External id": 531578,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622158932.249, "dur": 50.312,
    "args": {
      "External id": 531579,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622159002.208, "dur": 19.944,
    "args": {
      "External id": 531580,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5739
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.11)", "pid": 4183438, "tid": 4183438,
    "ts": 679622159200.441, "dur": 77.866,
    "args": {
      "External id": 531581,"Record function id": 0, "Ev Idx": 5740
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622159353.782, "dur": 48.491,
    "args": {
      "External id": 531582,"Record function id": 0, "Ev Idx": 5741
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.12)", "pid": 4183438, "tid": 4183438,
    "ts": 679622159411.933, "dur": 8844.238,
    "args": {
      "External id": 531583,"Record function id": 0, "Ev Idx": 5742
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.12)", "pid": 4183438, "tid": 4183438,
    "ts": 679622159420.831, "dur": 929.717,
    "args": {
      "External id": 531584,"Record function id": 0, "Ev Idx": 5743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622159500.838, "dur": 9.746,
    "args": {
      "External id": 531585,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622159526.646, "dur": 41.207,
    "args": {
      "External id": 531586,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159532.668, "dur": 2.564,
    "args": {
      "External id": 531587,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159539.917, "dur": 0.416,
    "args": {
      "External id": 531588,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159541.992, "dur": 0.570,
    "args": {
      "External id": 531589,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159544.371, "dur": 2.791,
    "args": {
      "External id": 531590,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159548.770, "dur": 0.334,
    "args": {
      "External id": 531591,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159550.561, "dur": 0.455,
    "args": {
      "External id": 531592,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159554.824, "dur": 1.462,
    "args": {
      "External id": 531593,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159558.158, "dur": 0.372,
    "args": {
      "External id": 531594,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159560.085, "dur": 0.292,
    "args": {
      "External id": 531595,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622159580.234, "dur": 43.254,
    "args": {
      "External id": 531596,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622159695.300, "dur": 126.386,
    "args": {
      "External id": 531597,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622159707.596, "dur": 6.490,
    "args": {
      "External id": 531598,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622159719.932, "dur": 13.574,
    "args": {
      "External id": 531599,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622159724.694, "dur": 8.391,
    "args": {
      "External id": 531600,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159728.434, "dur": 2.829,
    "args": {
      "External id": 531601,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622159741.223, "dur": 31.505,
    "args": {
      "External id": 531602,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159743.764, "dur": 0.566,
    "args": {
      "External id": 531603,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159746.503, "dur": 0.490,
    "args": {
      "External id": 531604,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159748.849, "dur": 0.135,
    "args": {
      "External id": 531605,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159752.528, "dur": 1.525,
    "args": {
      "External id": 531606,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159755.290, "dur": 0.324,
    "args": {
      "External id": 531607,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159757.187, "dur": 2.740,
    "args": {
      "External id": 531608,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159761.372, "dur": 0.329,
    "args": {
      "External id": 531609,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159763.382, "dur": 0.268,
    "args": {
      "External id": 531610,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622159767.385, "dur": 0.380,
    "args": {
      "External id": 531611,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622159784.558, "dur": 28.159,
    "args": {
      "External id": 531612,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622159877.705, "dur": 377.169,
    "args": {
      "External id": 531613,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622159905.599, "dur": 343.721,
    "args": {
      "External id": 531614,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5773, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622159916.048, "dur": 327.126,
    "args": {
      "External id": 531615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622160275.193, "dur": 2.420,
    "args": {
      "External id": 531616,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5775, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.12)", "pid": 4183438, "tid": 4183438,
    "ts": 679622160371.470, "dur": 7683.209,
    "args": {
      "External id": 531617,"Record function id": 0, "Ev Idx": 5776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160470.266, "dur": 6.754,
    "args": {
      "External id": 531618,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160480.641, "dur": 1.060,
    "args": {
      "External id": 531619,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160483.742, "dur": 1.866,
    "args": {
      "External id": 531620,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160487.369, "dur": 1.033,
    "args": {
      "External id": 531621,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160490.122, "dur": 0.852,
    "args": {
      "External id": 531622,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160492.501, "dur": 0.695,
    "args": {
      "External id": 531623,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160494.841, "dur": 0.831,
    "args": {
      "External id": 531624,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160497.164, "dur": 2.435,
    "args": {
      "External id": 531625,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160501.441, "dur": 1.283,
    "args": {
      "External id": 531626,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622160504.287, "dur": 0.660,
    "args": {
      "External id": 531627,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622160523.765, "dur": 7492.649,
    "args": {
      "External id": 531628,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622160540.173, "dur": 7469.604,
    "args": {
      "External id": 531629,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622160560.548, "dur": 14.539,
    "args": {
      "External id": 531630,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622160577.801, "dur": 7397.883,
    "args": {
      "External id": 531631,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622160580.618, "dur": 7394.487,
    "args": {
      "External id": 531632,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622160586.928, "dur": 6.919,
    "args": {
      "External id": 531633,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622160595.653, "dur": 7376.545,
    "args": {
      "External id": 531634,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622168196.757, "dur": 34.568,
    "args": {
      "External id": 531635,"Sequence number": 6319441, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5794
    }
  },
  {
    "ph": "s", "id": 187, "pid": 4183438, "tid": 4183438, "ts": 679622168196.757,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622168217.703, "dur": 8.289,
    "args": {
      "External id": 531636,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622168220.863, "dur": 4.763,
    "args": {
      "External id": 531637,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622168294.603, "dur": 89.479,
    "args": {
      "External id": 531638,"Record function id": 0, "Ev Idx": 5797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622168385.866, "dur": 1090.798,
    "args": {
      "External id": 531639,"Record function id": 0, "Ev Idx": 5798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622168425.218, "dur": 1037.458,
    "args": {
      "External id": 531640,"Sequence number": 6319442, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5799
    }
  },
  {
    "ph": "s", "id": 186, "pid": 4183438, "tid": 4183438, "ts": 679622168425.218,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622168491.960, "dur": 44.433,
    "args": {
      "External id": 531641,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622168549.072, "dur": 92.205,
    "args": {
      "External id": 531642,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622168650.285, "dur": 81.941,
    "args": {
      "External id": 531643,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622168746.127, "dur": 29.978,
    "args": {
      "External id": 531644,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622168801.718, "dur": 27.746,
    "args": {
      "External id": 531645,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622168848.086, "dur": 14.886,
    "args": {
      "External id": 531646,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622168881.187, "dur": 129.622,
    "args": {
      "External id": 531647,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622168932.575, "dur": 12.608,
    "args": {
      "External id": 531648,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622168938.347, "dur": 5.997,
    "args": {
      "External id": 531649,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622168947.676, "dur": 6.475,
    "args": {
      "External id": 531650,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622168955.494, "dur": 1.620,
    "args": {
      "External id": 531651,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622168959.418, "dur": 2.962,
    "args": {
      "External id": 531652,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622169023.021, "dur": 45.505,
    "args": {
      "External id": 531653,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622169114.094, "dur": 29.967,
    "args": {
      "External id": 531654,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622169154.982, "dur": 41.593,
    "args": {
      "External id": 531655,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622169205.791, "dur": 47.629,
    "args": {
      "External id": 531656,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622169282.194, "dur": 26.501,
    "args": {
      "External id": 531657,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622169316.685, "dur": 45.169,
    "args": {
      "External id": 531658,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622169377.952, "dur": 17.808,
    "args": {
      "External id": 531659,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5818
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.12)", "pid": 4183438, "tid": 4183438,
    "ts": 679622169541.999, "dur": 77.196,
    "args": {
      "External id": 531660,"Record function id": 0, "Ev Idx": 5819
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622169737.214, "dur": 53.090,
    "args": {
      "External id": 531661,"Record function id": 0, "Ev Idx": 5820
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.13)", "pid": 4183438, "tid": 4183438,
    "ts": 679622169800.167, "dur": 8699.918,
    "args": {
      "External id": 531662,"Record function id": 0, "Ev Idx": 5821
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.13)", "pid": 4183438, "tid": 4183438,
    "ts": 679622169810.640, "dur": 900.193,
    "args": {
      "External id": 531663,"Record function id": 0, "Ev Idx": 5822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622169892.047, "dur": 9.784,
    "args": {
      "External id": 531664,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622169915.108, "dur": 39.359,
    "args": {
      "External id": 531665,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169921.130, "dur": 2.212,
    "args": {
      "External id": 531666,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169927.746, "dur": 0.902,
    "args": {
      "External id": 531667,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169930.206, "dur": 0.784,
    "args": {
      "External id": 531668,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169932.442, "dur": 0.694,
    "args": {
      "External id": 531669,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169936.357, "dur": 0.608,
    "args": {
      "External id": 531670,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169938.338, "dur": 0.812,
    "args": {
      "External id": 531671,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169940.954, "dur": 2.475,
    "args": {
      "External id": 531672,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169945.013, "dur": 0.617,
    "args": {
      "External id": 531673,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622169947.107, "dur": 0.372,
    "args": {
      "External id": 531674,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622169964.712, "dur": 42.303,
    "args": {
      "External id": 531675,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622170039.691, "dur": 136.806,
    "args": {
      "External id": 531676,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622170050.001, "dur": 4.503,
    "args": {
      "External id": 531677,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622170059.839, "dur": 10.162,
    "args": {
      "External id": 531678,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622170064.271, "dur": 5.293,
    "args": {
      "External id": 531679,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170067.736, "dur": 0.764,
    "args": {
      "External id": 531680,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622170076.637, "dur": 50.595,
    "args": {
      "External id": 531681,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170078.506, "dur": 3.080,
    "args": {
      "External id": 531682,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170083.285, "dur": 0.801,
    "args": {
      "External id": 531683,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170085.734, "dur": 0.814,
    "args": {
      "External id": 531684,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170090.020, "dur": 0.637,
    "args": {
      "External id": 531685,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170091.911, "dur": 0.646,
    "args": {
      "External id": 531686,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170094.031, "dur": 0.727,
    "args": {
      "External id": 531687,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170114.887, "dur": 0.736,
    "args": {
      "External id": 531688,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170117.394, "dur": 0.497,
    "args": {
      "External id": 531689,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170119.359, "dur": 2.515,
    "args": {
      "External id": 531690,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622170142.034, "dur": 25.794,
    "args": {
      "External id": 531691,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622170229.925, "dur": 359.192,
    "args": {
      "External id": 531692,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622170256.161, "dur": 328.508,
    "args": {
      "External id": 531693,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5852, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622170266.818, "dur": 310.752,
    "args": {
      "External id": 531694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622170607.668, "dur": 2.236,
    "args": {
      "External id": 531695,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5854, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.13)", "pid": 4183438, "tid": 4183438,
    "ts": 679622170735.105, "dur": 7575.442,
    "args": {
      "External id": 531696,"Record function id": 0, "Ev Idx": 5855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170833.572, "dur": 6.664,
    "args": {
      "External id": 531697,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170843.644, "dur": 1.691,
    "args": {
      "External id": 531698,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170846.985, "dur": 1.623,
    "args": {
      "External id": 531699,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170850.516, "dur": 1.060,
    "args": {
      "External id": 531700,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170853.161, "dur": 1.193,
    "args": {
      "External id": 531701,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170856.432, "dur": 1.409,
    "args": {
      "External id": 531702,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170860.079, "dur": 0.977,
    "args": {
      "External id": 531703,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170863.552, "dur": 2.146,
    "args": {
      "External id": 531704,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170867.332, "dur": 1.199,
    "args": {
      "External id": 531705,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622170870.330, "dur": 0.956,
    "args": {
      "External id": 531706,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622170890.272, "dur": 7381.530,
    "args": {
      "External id": 531707,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622170905.368, "dur": 7359.532,
    "args": {
      "External id": 531708,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622170923.106, "dur": 14.817,
    "args": {
      "External id": 531709,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622170940.911, "dur": 7292.716,
    "args": {
      "External id": 531710,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622170943.610, "dur": 7289.547,
    "args": {
      "External id": 531711,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622170949.776, "dur": 7.122,
    "args": {
      "External id": 531712,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622170958.966, "dur": 7271.222,
    "args": {
      "External id": 531713,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622178446.482, "dur": 28.962,
    "args": {
      "External id": 531714,"Sequence number": 6319443, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5873
    }
  },
  {
    "ph": "s", "id": 185, "pid": 4183438, "tid": 4183438, "ts": 679622178446.482,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622178462.919, "dur": 7.821,
    "args": {
      "External id": 531715,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622178465.999, "dur": 4.498,
    "args": {
      "External id": 531716,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622178539.431, "dur": 88.716,
    "args": {
      "External id": 531717,"Record function id": 0, "Ev Idx": 5876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622178630.161, "dur": 1124.481,
    "args": {
      "External id": 531718,"Record function id": 0, "Ev Idx": 5877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622178717.833, "dur": 1022.246,
    "args": {
      "External id": 531719,"Sequence number": 6319444, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5878
    }
  },
  {
    "ph": "s", "id": 184, "pid": 4183438, "tid": 4183438, "ts": 679622178717.833,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622178787.927, "dur": 45.692,
    "args": {
      "External id": 531720,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622178847.662, "dur": 91.393,
    "args": {
      "External id": 531721,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622178948.024, "dur": 33.774,
    "args": {
      "External id": 531722,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622178991.232, "dur": 28.566,
    "args": {
      "External id": 531723,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622179044.570, "dur": 25.890,
    "args": {
      "External id": 531724,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622179087.981, "dur": 30.033,
    "args": {
      "External id": 531725,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622179138.551, "dur": 137.238,
    "args": {
      "External id": 531726,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622179191.961, "dur": 12.513,
    "args": {
      "External id": 531727,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622179197.179, "dur": 6.335,
    "args": {
      "External id": 531728,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622179207.300, "dur": 6.188,
    "args": {
      "External id": 531729,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622179215.218, "dur": 1.424,
    "args": {
      "External id": 531730,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622179219.032, "dur": 2.720,
    "args": {
      "External id": 531731,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622179287.462, "dur": 54.160,
    "args": {
      "External id": 531732,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622179372.182, "dur": 29.953,
    "args": {
      "External id": 531733,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622179410.328, "dur": 39.150,
    "args": {
      "External id": 531734,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622179458.058, "dur": 32.037,
    "args": {
      "External id": 531735,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622179512.190, "dur": 28.751,
    "args": {
      "External id": 531736,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622179546.944, "dur": 43.310,
    "args": {
      "External id": 531737,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622179609.332, "dur": 20.121,
    "args": {
      "External id": 531738,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5897
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.13)", "pid": 4183438, "tid": 4183438,
    "ts": 679622179822.910, "dur": 78.472,
    "args": {
      "External id": 531739,"Record function id": 0, "Ev Idx": 5898
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622179975.311, "dur": 49.198,
    "args": {
      "External id": 531740,"Record function id": 0, "Ev Idx": 5899
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.14)", "pid": 4183438, "tid": 4183438,
    "ts": 679622180033.151, "dur": 8943.377,
    "args": {
      "External id": 531741,"Record function id": 0, "Ev Idx": 5900
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.14)", "pid": 4183438, "tid": 4183438,
    "ts": 679622180041.531, "dur": 906.801,
    "args": {
      "External id": 531742,"Record function id": 0, "Ev Idx": 5901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622180138.311, "dur": 9.760,
    "args": {
      "External id": 531743,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622180161.562, "dur": 45.890,
    "args": {
      "External id": 531744,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180170.517, "dur": 3.170,
    "args": {
      "External id": 531745,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180178.012, "dur": 1.073,
    "args": {
      "External id": 531746,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180180.327, "dur": 0.847,
    "args": {
      "External id": 531747,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180182.439, "dur": 2.857,
    "args": {
      "External id": 531748,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180186.489, "dur": 0.402,
    "args": {
      "External id": 531749,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180191.061, "dur": 0.489,
    "args": {
      "External id": 531750,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180194.069, "dur": 0.591,
    "args": {
      "External id": 531751,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180196.269, "dur": 0.769,
    "args": {
      "External id": 531752,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180198.045, "dur": 2.390,
    "args": {
      "External id": 531753,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622180218.373, "dur": 40.461,
    "args": {
      "External id": 531754,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622180292.911, "dur": 106.089,
    "args": {
      "External id": 531755,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622180302.946, "dur": 4.436,
    "args": {
      "External id": 531756,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622180313.048, "dur": 10.738,
    "args": {
      "External id": 531757,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622180317.617, "dur": 5.759,
    "args": {
      "External id": 531758,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180321.358, "dur": 0.702,
    "args": {
      "External id": 531759,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622180330.588, "dur": 28.216,
    "args": {
      "External id": 531760,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180332.523, "dur": 0.561,
    "args": {
      "External id": 531761,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180335.252, "dur": 0.409,
    "args": {
      "External id": 531762,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180336.623, "dur": 2.447,
    "args": {
      "External id": 531763,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180340.509, "dur": 0.471,
    "args": {
      "External id": 531764,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180342.056, "dur": 0.415,
    "args": {
      "External id": 531765,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180346.629, "dur": 0.500,
    "args": {
      "External id": 531766,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180348.271, "dur": 0.655,
    "args": {
      "External id": 531767,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180350.398, "dur": 0.726,
    "args": {
      "External id": 531768,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622180353.940, "dur": 0.717,
    "args": {
      "External id": 531769,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622180369.770, "dur": 21.230,
    "args": {
      "External id": 531770,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622180454.098, "dur": 395.514,
    "args": {
      "External id": 531771,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 5930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622180480.857, "dur": 363.660,
    "args": {
      "External id": 531772,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5931, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622180491.288, "dur": 347.412,
    "args": {
      "External id": 531773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 5932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622180872.207, "dur": 2.321,
    "args": {
      "External id": 531774,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5933, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.14)", "pid": 4183438, "tid": 4183438,
    "ts": 679622180969.987, "dur": 7812.549,
    "args": {
      "External id": 531775,"Record function id": 0, "Ev Idx": 5934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181069.979, "dur": 7.119,
    "args": {
      "External id": 531776,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 5935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181080.854, "dur": 1.528,
    "args": {
      "External id": 531777,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181084.958, "dur": 1.164,
    "args": {
      "External id": 531778,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181087.815, "dur": 1.134,
    "args": {
      "External id": 531779,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181090.852, "dur": 0.955,
    "args": {
      "External id": 531780,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181094.038, "dur": 1.080,
    "args": {
      "External id": 531781,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 5940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181113.239, "dur": 1.804,
    "args": {
      "External id": 531782,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 5941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181118.895, "dur": 2.434,
    "args": {
      "External id": 531783,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181123.363, "dur": 0.940,
    "args": {
      "External id": 531784,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622181126.183, "dur": 1.056,
    "args": {
      "External id": 531785,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 5944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622181147.210, "dur": 7596.766,
    "args": {
      "External id": 531786,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622181163.135, "dur": 7573.836,
    "args": {
      "External id": 531787,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 5946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622181179.321, "dur": 15.264,
    "args": {
      "External id": 531788,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622181197.312, "dur": 7507.166,
    "args": {
      "External id": 531789,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 5948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622181199.938, "dur": 7504.020,
    "args": {
      "External id": 531790,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 5949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622181206.238, "dur": 7.626,
    "args": {
      "External id": 531791,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622181215.957, "dur": 7484.971,
    "args": {
      "External id": 531792,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 5951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622188921.406, "dur": 29.948,
    "args": {
      "External id": 531793,"Sequence number": 6319445, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 5952
    }
  },
  {
    "ph": "s", "id": 183, "pid": 4183438, "tid": 4183438, "ts": 679622188921.406,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622188938.515, "dur": 7.958,
    "args": {
      "External id": 531794,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 5953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622188941.890, "dur": 4.367,
    "args": {
      "External id": 531795,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 5954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622189015.156, "dur": 105.158,
    "args": {
      "External id": 531796,"Record function id": 0, "Ev Idx": 5955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622189123.707, "dur": 1122.926,
    "args": {
      "External id": 531797,"Record function id": 0, "Ev Idx": 5956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622189165.835, "dur": 1066.653,
    "args": {
      "External id": 531798,"Sequence number": 6319446, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 5957
    }
  },
  {
    "ph": "s", "id": 182, "pid": 4183438, "tid": 4183438, "ts": 679622189165.835,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622189234.670, "dur": 46.847,
    "args": {
      "External id": 531799,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622189295.004, "dur": 92.160,
    "args": {
      "External id": 531800,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622189396.430, "dur": 34.994,
    "args": {
      "External id": 531801,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622189440.604, "dur": 29.302,
    "args": {
      "External id": 531802,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622189496.516, "dur": 28.213,
    "args": {
      "External id": 531803,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622189543.217, "dur": 16.020,
    "args": {
      "External id": 531804,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622189576.124, "dur": 185.351,
    "args": {
      "External id": 531805,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 5964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622189634.176, "dur": 11.147,
    "args": {
      "External id": 531806,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 5965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622189638.869, "dur": 5.571,
    "args": {
      "External id": 531807,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622189648.110, "dur": 46.441,
    "args": {
      "External id": 531808,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622189697.124, "dur": 1.893,
    "args": {
      "External id": 531809,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622189701.499, "dur": 3.236,
    "args": {
      "External id": 531810,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622189774.067, "dur": 54.813,
    "args": {
      "External id": 531811,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 5970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622189862.529, "dur": 30.007,
    "args": {
      "External id": 531812,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 5971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622189901.049, "dur": 38.669,
    "args": {
      "External id": 531813,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622189949.103, "dur": 50.581,
    "args": {
      "External id": 531814,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 5973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622190021.819, "dur": 26.218,
    "args": {
      "External id": 531815,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 5974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622190054.266, "dur": 60.264,
    "args": {
      "External id": 531816,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 5975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622190138.679, "dur": 23.427,
    "args": {
      "External id": 531817,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 5976
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.14)", "pid": 4183438, "tid": 4183438,
    "ts": 679622190312.039, "dur": 77.362,
    "args": {
      "External id": 531818,"Record function id": 0, "Ev Idx": 5977
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622190462.069, "dur": 49.133,
    "args": {
      "External id": 531819,"Record function id": 0, "Ev Idx": 5978
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.15)", "pid": 4183438, "tid": 4183438,
    "ts": 679622190520.005, "dur": 9254.965,
    "args": {
      "External id": 531820,"Record function id": 0, "Ev Idx": 5979
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.15)", "pid": 4183438, "tid": 4183438,
    "ts": 679622190528.308, "dur": 995.201,
    "args": {
      "External id": 531821,"Record function id": 0, "Ev Idx": 5980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622190609.424, "dur": 9.251,
    "args": {
      "External id": 531822,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622190631.604, "dur": 81.393,
    "args": {
      "External id": 531823,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190637.404, "dur": 2.278,
    "args": {
      "External id": 531824,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190643.982, "dur": 0.694,
    "args": {
      "External id": 531825,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190646.545, "dur": 0.862,
    "args": {
      "External id": 531826,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190648.217, "dur": 0.887,
    "args": {
      "External id": 531827,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190694.434, "dur": 0.687,
    "args": {
      "External id": 531828,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190697.180, "dur": 0.802,
    "args": {
      "External id": 531829,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190699.385, "dur": 3.126,
    "args": {
      "External id": 531830,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190703.549, "dur": 0.662,
    "args": {
      "External id": 531831,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190705.529, "dur": 0.667,
    "args": {
      "External id": 531832,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622190724.776, "dur": 46.304,
    "args": {
      "External id": 531833,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 5992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622190807.354, "dur": 129.038,
    "args": {
      "External id": 531834,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 5993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622190818.935, "dur": 6.479,
    "args": {
      "External id": 531835,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622190831.131, "dur": 10.995,
    "args": {
      "External id": 531836,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622190835.681, "dur": 6.027,
    "args": {
      "External id": 531837,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 5996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190839.741, "dur": 0.707,
    "args": {
      "External id": 531838,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 5997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622190849.291, "dur": 32.459,
    "args": {
      "External id": 531839,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 5998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190851.309, "dur": 2.822,
    "args": {
      "External id": 531840,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 5999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190855.472, "dur": 0.889,
    "args": {
      "External id": 531841,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190858.286, "dur": 1.046,
    "args": {
      "External id": 531842,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190862.466, "dur": 0.755,
    "args": {
      "External id": 531843,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190864.535, "dur": 0.806,
    "args": {
      "External id": 531844,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190866.581, "dur": 0.386,
    "args": {
      "External id": 531845,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190870.980, "dur": 0.739,
    "args": {
      "External id": 531846,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190872.811, "dur": 0.668,
    "args": {
      "External id": 531847,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622190874.812, "dur": 2.654,
    "args": {
      "External id": 531848,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622190901.574, "dur": 25.363,
    "args": {
      "External id": 531849,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622190990.759, "dur": 434.746,
    "args": {
      "External id": 531850,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 6009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622191017.722, "dur": 401.740,
    "args": {
      "External id": 531851,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6010, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622191030.103, "dur": 383.132,
    "args": {
      "External id": 531852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 6011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622191448.300, "dur": 2.779,
    "args": {
      "External id": 531853,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6012, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.15)", "pid": 4183438, "tid": 4183438,
    "ts": 679622191545.256, "dur": 8007.570,
    "args": {
      "External id": 531854,"Record function id": 0, "Ev Idx": 6013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191640.327, "dur": 6.907,
    "args": {
      "External id": 531855,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 6014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191688.242, "dur": 2.201,
    "args": {
      "External id": 531856,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191694.305, "dur": 1.095,
    "args": {
      "External id": 531857,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191697.444, "dur": 1.010,
    "args": {
      "External id": 531858,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191700.112, "dur": 1.438,
    "args": {
      "External id": 531859,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191703.605, "dur": 1.101,
    "args": {
      "External id": 531860,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191706.509, "dur": 1.179,
    "args": {
      "External id": 531861,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191710.072, "dur": 2.641,
    "args": {
      "External id": 531862,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191714.630, "dur": 1.301,
    "args": {
      "External id": 531863,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622191717.641, "dur": 1.180,
    "args": {
      "External id": 531864,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622191737.956, "dur": 7776.500,
    "args": {
      "External id": 531865,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622191754.178, "dur": 7752.890,
    "args": {
      "External id": 531866,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622191770.945, "dur": 14.701,
    "args": {
      "External id": 531867,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622191788.550, "dur": 7686.159,
    "args": {
      "External id": 531868,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 6027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622191791.494, "dur": 7682.765,
    "args": {
      "External id": 531869,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 6028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622191797.908, "dur": 7.901,
    "args": {
      "External id": 531870,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622191807.594, "dur": 7663.425,
    "args": {
      "External id": 531871,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 6030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622199716.164, "dur": 31.604,
    "args": {
      "External id": 531872,"Sequence number": 6319447, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 6031
    }
  },
  {
    "ph": "s", "id": 181, "pid": 4183438, "tid": 4183438, "ts": 679622199716.164,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622199733.728, "dur": 9.000,
    "args": {
      "External id": 531873,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 6032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622199737.416, "dur": 4.979,
    "args": {
      "External id": 531874,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622199812.274, "dur": 91.444,
    "args": {
      "External id": 531875,"Record function id": 0, "Ev Idx": 6034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622199905.776, "dur": 1105.722,
    "args": {
      "External id": 531876,"Record function id": 0, "Ev Idx": 6035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622199950.654, "dur": 1047.040,
    "args": {
      "External id": 531877,"Sequence number": 6319448, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 6036
    }
  },
  {
    "ph": "s", "id": 180, "pid": 4183438, "tid": 4183438, "ts": 679622199950.654,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622200018.586, "dur": 44.653,
    "args": {
      "External id": 531878,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622200076.381, "dur": 113.063,
    "args": {
      "External id": 531879,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622200201.831, "dur": 37.906,
    "args": {
      "External id": 531880,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622200248.727, "dur": 30.226,
    "args": {
      "External id": 531881,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622200304.658, "dur": 27.327,
    "args": {
      "External id": 531882,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622200350.860, "dur": 17.394,
    "args": {
      "External id": 531883,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622200386.322, "dur": 131.871,
    "args": {
      "External id": 531884,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622200436.925, "dur": 11.933,
    "args": {
      "External id": 531885,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622200441.830, "dur": 6.175,
    "args": {
      "External id": 531886,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622200451.485, "dur": 6.108,
    "args": {
      "External id": 531887,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622200459.522, "dur": 1.809,
    "args": {
      "External id": 531888,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622200463.626, "dur": 3.019,
    "args": {
      "External id": 531889,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622200530.669, "dur": 42.148,
    "args": {
      "External id": 531890,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622200602.530, "dur": 30.214,
    "args": {
      "External id": 531891,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622200641.733, "dur": 79.640,
    "args": {
      "External id": 531892,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622200734.901, "dur": 39.845,
    "args": {
      "External id": 531893,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622200797.660, "dur": 29.265,
    "args": {
      "External id": 531894,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 6053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622200833.097, "dur": 51.402,
    "args": {
      "External id": 531895,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 6054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622200904.380, "dur": 18.265,
    "args": {
      "External id": 531896,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 6055
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.15)", "pid": 4183438, "tid": 4183438,
    "ts": 679622201076.103, "dur": 99.098,
    "args": {
      "External id": 531897,"Record function id": 0, "Ev Idx": 6056
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622201250.495, "dur": 50.872,
    "args": {
      "External id": 531898,"Record function id": 0, "Ev Idx": 6057
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.16)", "pid": 4183438, "tid": 4183438,
    "ts": 679622201309.880, "dur": 9155.764,
    "args": {
      "External id": 531899,"Record function id": 0, "Ev Idx": 6058
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.16)", "pid": 4183438, "tid": 4183438,
    "ts": 679622201320.216, "dur": 916.929,
    "args": {
      "External id": 531900,"Record function id": 0, "Ev Idx": 6059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622201401.772, "dur": 9.529,
    "args": {
      "External id": 531901,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622201424.345, "dur": 37.524,
    "args": {
      "External id": 531902,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201429.771, "dur": 2.389,
    "args": {
      "External id": 531903,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201437.004, "dur": 0.590,
    "args": {
      "External id": 531904,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201438.994, "dur": 0.697,
    "args": {
      "External id": 531905,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201440.870, "dur": 0.744,
    "args": {
      "External id": 531906,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201444.334, "dur": 0.680,
    "args": {
      "External id": 531907,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201446.741, "dur": 0.726,
    "args": {
      "External id": 531908,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201448.478, "dur": 2.572,
    "args": {
      "External id": 531909,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201453.146, "dur": 0.461,
    "args": {
      "External id": 531910,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201454.833, "dur": 0.766,
    "args": {
      "External id": 531911,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622201475.525, "dur": 42.557,
    "args": {
      "External id": 531912,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622201551.356, "dur": 151.751,
    "args": {
      "External id": 531913,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 6072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622201561.697, "dur": 5.601,
    "args": {
      "External id": 531914,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622201572.868, "dur": 10.600,
    "args": {
      "External id": 531915,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622201577.550, "dur": 5.355,
    "args": {
      "External id": 531916,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 6075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201581.023, "dur": 0.655,
    "args": {
      "External id": 531917,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622201589.877, "dur": 31.075,
    "args": {
      "External id": 531918,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201591.768, "dur": 0.835,
    "args": {
      "External id": 531919,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201594.090, "dur": 2.627,
    "args": {
      "External id": 531920,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201598.017, "dur": 0.682,
    "args": {
      "External id": 531921,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201600.413, "dur": 0.834,
    "args": {
      "External id": 531922,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201605.451, "dur": 0.578,
    "args": {
      "External id": 531923,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201608.067, "dur": 0.869,
    "args": {
      "External id": 531924,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201609.970, "dur": 0.679,
    "args": {
      "External id": 531925,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201614.618, "dur": 0.418,
    "args": {
      "External id": 531926,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622201616.019, "dur": 0.417,
    "args": {
      "External id": 531927,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622201630.876, "dur": 62.465,
    "args": {
      "External id": 531928,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622201760.118, "dur": 382.209,
    "args": {
      "External id": 531929,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 6088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622201786.641, "dur": 350.324,
    "args": {
      "External id": 531930,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6089, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622201797.397, "dur": 333.608,
    "args": {
      "External id": 531931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 6090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622202164.408, "dur": 2.644,
    "args": {
      "External id": 531932,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6091, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.16)", "pid": 4183438, "tid": 4183438,
    "ts": 679622202258.175, "dur": 8015.176,
    "args": {
      "External id": 531933,"Record function id": 0, "Ev Idx": 6092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202355.800, "dur": 7.291,
    "args": {
      "External id": 531934,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 6093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202366.269, "dur": 1.445,
    "args": {
      "External id": 531935,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202369.970, "dur": 1.153,
    "args": {
      "External id": 531936,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202372.616, "dur": 1.287,
    "args": {
      "External id": 531937,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202375.724, "dur": 1.366,
    "args": {
      "External id": 531938,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202378.516, "dur": 1.384,
    "args": {
      "External id": 531939,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202382.098, "dur": 1.308,
    "args": {
      "External id": 531940,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202385.291, "dur": 2.395,
    "args": {
      "External id": 531941,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202389.542, "dur": 1.123,
    "args": {
      "External id": 531942,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622202392.464, "dur": 1.250,
    "args": {
      "External id": 531943,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622202411.616, "dur": 7823.217,
    "args": {
      "External id": 531944,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622202426.906, "dur": 7800.923,
    "args": {
      "External id": 531945,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622202443.296, "dur": 14.772,
    "args": {
      "External id": 531946,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622202461.042, "dur": 7734.914,
    "args": {
      "External id": 531947,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 6106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622202463.995, "dur": 7731.540,
    "args": {
      "External id": 531948,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 6107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622202470.752, "dur": 6.716,
    "args": {
      "External id": 531949,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622202479.340, "dur": 7713.167,
    "args": {
      "External id": 531950,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 6109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622210410.030, "dur": 30.671,
    "args": {
      "External id": 531951,"Sequence number": 6319449, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 6110
    }
  },
  {
    "ph": "s", "id": 179, "pid": 4183438, "tid": 4183438, "ts": 679622210410.030,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622210427.356, "dur": 8.431,
    "args": {
      "External id": 531952,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 6111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622210430.553, "dur": 4.977,
    "args": {
      "External id": 531953,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622210503.079, "dur": 88.349,
    "args": {
      "External id": 531954,"Record function id": 0, "Ev Idx": 6113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622210593.069, "dur": 1169.741,
    "args": {
      "External id": 531955,"Record function id": 0, "Ev Idx": 6114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622210631.554, "dur": 1115.610,
    "args": {
      "External id": 531956,"Sequence number": 6319450, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 6115
    }
  },
  {
    "ph": "s", "id": 178, "pid": 4183438, "tid": 4183438, "ts": 679622210631.554,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622210733.383, "dur": 48.477,
    "args": {
      "External id": 531957,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622210796.138, "dur": 93.501,
    "args": {
      "External id": 531958,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622210898.662, "dur": 34.682,
    "args": {
      "External id": 531959,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622210942.382, "dur": 29.384,
    "args": {
      "External id": 531960,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622211001.773, "dur": 27.809,
    "args": {
      "External id": 531961,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622211044.988, "dur": 16.527,
    "args": {
      "External id": 531962,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622211078.315, "dur": 156.128,
    "args": {
      "External id": 531963,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622211147.874, "dur": 13.146,
    "args": {
      "External id": 531964,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622211153.197, "dur": 6.678,
    "args": {
      "External id": 531965,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622211164.348, "dur": 5.866,
    "args": {
      "External id": 531966,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622211171.322, "dur": 1.941,
    "args": {
      "External id": 531967,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622211175.983, "dur": 4.668,
    "args": {
      "External id": 531968,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622211274.080, "dur": 60.952,
    "args": {
      "External id": 531969,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622211368.670, "dur": 27.300,
    "args": {
      "External id": 531970,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622211404.631, "dur": 38.078,
    "args": {
      "External id": 531971,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622211451.003, "dur": 45.779,
    "args": {
      "External id": 531972,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622211518.214, "dur": 24.559,
    "args": {
      "External id": 531973,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 6132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622211548.441, "dur": 55.978,
    "args": {
      "External id": 531974,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 6133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622211623.137, "dur": 18.276,
    "args": {
      "External id": 531975,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 6134
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.16)", "pid": 4183438, "tid": 4183438,
    "ts": 679622211828.633, "dur": 76.510,
    "args": {
      "External id": 531976,"Record function id": 0, "Ev Idx": 6135
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622211974.935, "dur": 48.131,
    "args": {
      "External id": 531977,"Record function id": 0, "Ev Idx": 6136
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.17)", "pid": 4183438, "tid": 4183438,
    "ts": 679622212031.943, "dur": 8986.219,
    "args": {
      "External id": 531978,"Record function id": 0, "Ev Idx": 6137
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.17)", "pid": 4183438, "tid": 4183438,
    "ts": 679622212040.163, "dur": 912.071,
    "args": {
      "External id": 531979,"Record function id": 0, "Ev Idx": 6138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622212137.825, "dur": 10.590,
    "args": {
      "External id": 531980,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622212162.456, "dur": 36.167,
    "args": {
      "External id": 531981,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212167.954, "dur": 2.289,
    "args": {
      "External id": 531982,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212175.009, "dur": 0.603,
    "args": {
      "External id": 531983,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212176.718, "dur": 0.577,
    "args": {
      "External id": 531984,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212178.881, "dur": 0.464,
    "args": {
      "External id": 531985,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212182.549, "dur": 0.406,
    "args": {
      "External id": 531986,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212184.222, "dur": 0.503,
    "args": {
      "External id": 531987,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212185.893, "dur": 2.717,
    "args": {
      "External id": 531988,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212190.116, "dur": 0.375,
    "args": {
      "External id": 531989,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212191.392, "dur": 0.488,
    "args": {
      "External id": 531990,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622212209.955, "dur": 42.555,
    "args": {
      "External id": 531991,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622212285.862, "dur": 109.035,
    "args": {
      "External id": 531992,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 6151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622212296.381, "dur": 4.731,
    "args": {
      "External id": 531993,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622212306.465, "dur": 10.474,
    "args": {
      "External id": 531994,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622212311.133, "dur": 5.385,
    "args": {
      "External id": 531995,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 6154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212314.563, "dur": 0.622,
    "args": {
      "External id": 531996,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622212323.587, "dur": 29.506,
    "args": {
      "External id": 531997,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212325.802, "dur": 2.769,
    "args": {
      "External id": 531998,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212329.985, "dur": 0.428,
    "args": {
      "External id": 531999,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212331.668, "dur": 0.588,
    "args": {
      "External id": 532000,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212336.960, "dur": 0.283,
    "args": {
      "External id": 532001,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212338.353, "dur": 0.308,
    "args": {
      "External id": 532002,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212339.828, "dur": 0.440,
    "args": {
      "External id": 532003,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212342.880, "dur": 0.435,
    "args": {
      "External id": 532004,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212344.793, "dur": 0.465,
    "args": {
      "External id": 532005,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622212346.708, "dur": 2.283,
    "args": {
      "External id": 532006,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622212365.565, "dur": 21.241,
    "args": {
      "External id": 532007,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622212446.950, "dur": 404.659,
    "args": {
      "External id": 532008,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 6167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622212473.030, "dur": 373.282,
    "args": {
      "External id": 532009,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6168, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622212483.959, "dur": 355.615,
    "args": {
      "External id": 532010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 6169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622212874.467, "dur": 2.487,
    "args": {
      "External id": 532011,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6170, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.17)", "pid": 4183438, "tid": 4183438,
    "ts": 679622212974.612, "dur": 7849.609,
    "args": {
      "External id": 532012,"Record function id": 0, "Ev Idx": 6171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213073.431, "dur": 6.884,
    "args": {
      "External id": 532013,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 6172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213083.682, "dur": 1.194,
    "args": {
      "External id": 532014,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213086.683, "dur": 0.954,
    "args": {
      "External id": 532015,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213089.142, "dur": 1.079,
    "args": {
      "External id": 532016,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213092.062, "dur": 1.348,
    "args": {
      "External id": 532017,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213095.013, "dur": 1.121,
    "args": {
      "External id": 532018,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213115.108, "dur": 1.690,
    "args": {
      "External id": 532019,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213120.070, "dur": 2.015,
    "args": {
      "External id": 532020,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213123.817, "dur": 0.973,
    "args": {
      "External id": 532021,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622213126.631, "dur": 0.932,
    "args": {
      "External id": 532022,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622213147.017, "dur": 7637.688,
    "args": {
      "External id": 532023,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622213163.975, "dur": 7613.652,
    "args": {
      "External id": 532024,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622213181.731, "dur": 14.587,
    "args": {
      "External id": 532025,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622213199.073, "dur": 7546.158,
    "args": {
      "External id": 532026,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 6185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622213201.566, "dur": 7543.138,
    "args": {
      "External id": 532027,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 6186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622213208.055, "dur": 7.411,
    "args": {
      "External id": 532028,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622213217.367, "dur": 7524.169,
    "args": {
      "External id": 532029,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 6188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622220963.186, "dur": 29.061,
    "args": {
      "External id": 532030,"Sequence number": 6319451, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 6189
    }
  },
  {
    "ph": "s", "id": 177, "pid": 4183438, "tid": 4183438, "ts": 679622220963.186,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622220978.818, "dur": 8.092,
    "args": {
      "External id": 532031,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 6190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622220982.250, "dur": 4.424,
    "args": {
      "External id": 532032,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622221056.713, "dur": 102.176,
    "args": {
      "External id": 532033,"Record function id": 0, "Ev Idx": 6192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622221161.773, "dur": 1088.239,
    "args": {
      "External id": 532034,"Record function id": 0, "Ev Idx": 6193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622221204.320, "dur": 1031.199,
    "args": {
      "External id": 532035,"Sequence number": 6319452, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 6194
    }
  },
  {
    "ph": "s", "id": 176, "pid": 4183438, "tid": 4183438, "ts": 679622221204.320,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622221272.770, "dur": 42.621,
    "args": {
      "External id": 532036,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622221328.953, "dur": 91.449,
    "args": {
      "External id": 532037,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622221429.575, "dur": 34.692,
    "args": {
      "External id": 532038,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622221473.182, "dur": 29.307,
    "args": {
      "External id": 532039,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622221527.044, "dur": 28.603,
    "args": {
      "External id": 532040,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622221574.568, "dur": 16.289,
    "args": {
      "External id": 532041,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622221608.588, "dur": 175.595,
    "args": {
      "External id": 532042,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622221695.477, "dur": 13.595,
    "args": {
      "External id": 532043,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622221700.468, "dur": 7.612,
    "args": {
      "External id": 532044,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622221711.927, "dur": 6.708,
    "args": {
      "External id": 532045,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622221720.083, "dur": 1.699,
    "args": {
      "External id": 532046,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622221724.311, "dur": 2.599,
    "args": {
      "External id": 532047,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622221795.480, "dur": 53.422,
    "args": {
      "External id": 532048,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622221880.577, "dur": 30.619,
    "args": {
      "External id": 532049,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622221919.327, "dur": 38.108,
    "args": {
      "External id": 532050,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622221967.129, "dur": 35.704,
    "args": {
      "External id": 532051,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622222024.659, "dur": 26.594,
    "args": {
      "External id": 532052,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 6211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622222057.569, "dur": 58.857,
    "args": {
      "External id": 532053,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 6212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622222141.283, "dur": 19.925,
    "args": {
      "External id": 532054,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 6213
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.17)", "pid": 4183438, "tid": 4183438,
    "ts": 679622222320.067, "dur": 79.439,
    "args": {
      "External id": 532055,"Record function id": 0, "Ev Idx": 6214
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622222474.873, "dur": 49.992,
    "args": {
      "External id": 532056,"Record function id": 0, "Ev Idx": 6215
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.18)", "pid": 4183438, "tid": 4183438,
    "ts": 679622222534.470, "dur": 9075.334,
    "args": {
      "External id": 532057,"Record function id": 0, "Ev Idx": 6216
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.18)", "pid": 4183438, "tid": 4183438,
    "ts": 679622222543.068, "dur": 906.519,
    "args": {
      "External id": 532058,"Record function id": 0, "Ev Idx": 6217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622222622.509, "dur": 9.466,
    "args": {
      "External id": 532059,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622222645.669, "dur": 77.764,
    "args": {
      "External id": 532060,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222651.031, "dur": 41.307,
    "args": {
      "External id": 532061,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222698.429, "dur": 0.911,
    "args": {
      "External id": 532062,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222700.343, "dur": 0.804,
    "args": {
      "External id": 532063,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222702.507, "dur": 0.774,
    "args": {
      "External id": 532064,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222706.975, "dur": 0.603,
    "args": {
      "External id": 532065,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222709.012, "dur": 0.874,
    "args": {
      "External id": 532066,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222710.987, "dur": 2.068,
    "args": {
      "External id": 532067,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222714.141, "dur": 0.458,
    "args": {
      "External id": 532068,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222715.647, "dur": 0.989,
    "args": {
      "External id": 532069,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622222736.462, "dur": 42.670,
    "args": {
      "External id": 532070,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622222816.765, "dur": 115.267,
    "args": {
      "External id": 532071,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 6230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622222831.913, "dur": 6.869,
    "args": {
      "External id": 532072,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622222844.238, "dur": 10.121,
    "args": {
      "External id": 532073,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622222848.741, "dur": 5.165,
    "args": {
      "External id": 532074,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 6233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222852.011, "dur": 0.726,
    "args": {
      "External id": 532075,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622222861.839, "dur": 28.838,
    "args": {
      "External id": 532076,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222864.019, "dur": 0.529,
    "args": {
      "External id": 532077,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222865.972, "dur": 2.384,
    "args": {
      "External id": 532078,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222869.520, "dur": 0.713,
    "args": {
      "External id": 532079,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222871.577, "dur": 0.752,
    "args": {
      "External id": 532080,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222875.879, "dur": 0.718,
    "args": {
      "External id": 532081,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222877.601, "dur": 0.741,
    "args": {
      "External id": 532082,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222879.877, "dur": 0.850,
    "args": {
      "External id": 532083,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222883.766, "dur": 0.565,
    "args": {
      "External id": 532084,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622222885.753, "dur": 0.959,
    "args": {
      "External id": 532085,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622222902.066, "dur": 21.437,
    "args": {
      "External id": 532086,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622222988.780, "dur": 363.053,
    "args": {
      "External id": 532087,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 6246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622223014.885, "dur": 331.417,
    "args": {
      "External id": 532088,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6247, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622223025.526, "dur": 314.957,
    "args": {
      "External id": 532089,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 6248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622223374.269, "dur": 2.152,
    "args": {
      "External id": 532090,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6249, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.18)", "pid": 4183438, "tid": 4183438,
    "ts": 679622223471.130, "dur": 7940.747,
    "args": {
      "External id": 532091,"Record function id": 0, "Ev Idx": 6250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223571.615, "dur": 6.865,
    "args": {
      "External id": 532092,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 6251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223582.026, "dur": 1.703,
    "args": {
      "External id": 532093,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223585.653, "dur": 1.358,
    "args": {
      "External id": 532094,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223588.568, "dur": 1.217,
    "args": {
      "External id": 532095,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223591.491, "dur": 1.248,
    "args": {
      "External id": 532096,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223595.132, "dur": 0.899,
    "args": {
      "External id": 532097,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223597.867, "dur": 1.173,
    "args": {
      "External id": 532098,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223601.010, "dur": 2.836,
    "args": {
      "External id": 532099,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223605.533, "dur": 0.952,
    "args": {
      "External id": 532100,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622223608.472, "dur": 1.039,
    "args": {
      "External id": 532101,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622223627.660, "dur": 7744.732,
    "args": {
      "External id": 532102,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622223643.159, "dur": 7721.924,
    "args": {
      "External id": 532103,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622223696.766, "dur": 14.676,
    "args": {
      "External id": 532104,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622223714.588, "dur": 7616.442,
    "args": {
      "External id": 532105,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 6264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622223717.195, "dur": 7613.388,
    "args": {
      "External id": 532106,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 6265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622223723.774, "dur": 6.824,
    "args": {
      "External id": 532107,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622223732.411, "dur": 7595.089,
    "args": {
      "External id": 532108,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 6267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622231555.667, "dur": 29.420,
    "args": {
      "External id": 532109,"Sequence number": 6319453, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 6268
    }
  },
  {
    "ph": "s", "id": 175, "pid": 4183438, "tid": 4183438, "ts": 679622231555.667,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622231571.844, "dur": 8.353,
    "args": {
      "External id": 532110,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 6269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622231575.201, "dur": 4.756,
    "args": {
      "External id": 532111,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622231646.470, "dur": 117.443,
    "args": {
      "External id": 532112,"Record function id": 0, "Ev Idx": 6271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622231767.366, "dur": 1066.974,
    "args": {
      "External id": 532113,"Record function id": 0, "Ev Idx": 6272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622231809.186, "dur": 1010.140,
    "args": {
      "External id": 532114,"Sequence number": 6319454, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 6273
    }
  },
  {
    "ph": "s", "id": 174, "pid": 4183438, "tid": 4183438, "ts": 679622231809.186,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622231882.983, "dur": 45.200,
    "args": {
      "External id": 532115,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622231945.663, "dur": 94.214,
    "args": {
      "External id": 532116,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622232048.977, "dur": 34.704,
    "args": {
      "External id": 532117,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622232091.625, "dur": 47.559,
    "args": {
      "External id": 532118,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622232170.059, "dur": 25.806,
    "args": {
      "External id": 532119,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622232211.541, "dur": 17.409,
    "args": {
      "External id": 532120,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622232245.339, "dur": 129.684,
    "args": {
      "External id": 532121,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622232296.127, "dur": 12.669,
    "args": {
      "External id": 532122,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622232301.253, "dur": 6.631,
    "args": {
      "External id": 532123,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622232311.594, "dur": 4.574,
    "args": {
      "External id": 532124,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622232317.746, "dur": 3.597,
    "args": {
      "External id": 532125,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622232323.623, "dur": 3.048,
    "args": {
      "External id": 532126,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622232385.796, "dur": 45.483,
    "args": {
      "External id": 532127,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622232462.082, "dur": 27.478,
    "args": {
      "External id": 532128,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622232497.534, "dur": 38.580,
    "args": {
      "External id": 532129,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622232543.002, "dur": 31.743,
    "args": {
      "External id": 532130,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622232599.103, "dur": 24.141,
    "args": {
      "External id": 532131,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 6290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622232629.044, "dur": 76.941,
    "args": {
      "External id": 532132,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 6291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622232728.468, "dur": 22.451,
    "args": {
      "External id": 532133,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 6292
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.18)", "pid": 4183438, "tid": 4183438,
    "ts": 679622232899.460, "dur": 78.682,
    "args": {
      "External id": 532134,"Record function id": 0, "Ev Idx": 6293
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 4183438, "tid": 4183438,
    "ts": 679622233050.961, "dur": 64.785,
    "args": {
      "External id": 532135,"Record function id": 0, "Ev Idx": 6294
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.19)", "pid": 4183438, "tid": 4183438,
    "ts": 679622233127.946, "dur": 8811.518,
    "args": {
      "External id": 532136,"Record function id": 0, "Ev Idx": 6295
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.19)", "pid": 4183438, "tid": 4183438,
    "ts": 679622233137.486, "dur": 900.758,
    "args": {
      "External id": 532137,"Record function id": 0, "Ev Idx": 6296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622233221.462, "dur": 10.168,
    "args": {
      "External id": 532138,"Record function id": 0, "Concrete Inputs": ["[1605888]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622233244.963, "dur": 36.693,
    "args": {
      "External id": 532139,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233250.471, "dur": 2.206,
    "args": {
      "External id": 532140,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233257.537, "dur": 0.566,
    "args": {
      "External id": 532141,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "128"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233259.500, "dur": 0.797,
    "args": {
      "External id": 532142,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "131200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233261.442, "dur": 0.881,
    "args": {
      "External id": 532143,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "262272"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233265.354, "dur": 0.763,
    "args": {
      "External id": 532144,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "393344"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233267.643, "dur": 0.728,
    "args": {
      "External id": 532145,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "524416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233269.290, "dur": 2.646,
    "args": {
      "External id": 532146,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233273.455, "dur": 0.533,
    "args": {
      "External id": 532147,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "884992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233275.068, "dur": 0.576,
    "args": {
      "External id": 532148,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "1245440"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622233292.810, "dur": 41.734,
    "args": {
      "External id": 532149,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 4183438, "tid": 4183438,
    "ts": 679622233367.814, "dur": 111.017,
    "args": {
      "External id": 532150,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1605888", "8", "2", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [], [], [], [], [], []], "Ev Idx": 6309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622233378.247, "dur": 4.270,
    "args": {
      "External id": 532151,"Record function id": 0, "Concrete Inputs": ["[12847104]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 4183438, "tid": 4183438,
    "ts": 679622233387.784, "dur": 10.716,
    "args": {
      "External id": 532152,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "1605888"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622233392.361, "dur": 5.705,
    "args": {
      "External id": 532153,"Record function id": 0, "Concrete Inputs": ["", "0", "3211776", "4817664", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[12847104], [], [], [], []], "Ev Idx": 6312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233396.088, "dur": 0.818,
    "args": {
      "External id": 532154,"Record function id": 0, "Concrete Inputs": ["", "[1605888]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[12847104], [], [], []], "Ev Idx": 6313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 4183438, "tid": 4183438,
    "ts": 679622233405.294, "dur": 31.488,
    "args": {
      "External id": 532155,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1605888], [], []], "Ev Idx": 6314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233407.232, "dur": 3.064,
    "args": {
      "External id": 532156,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3211776"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233411.862, "dur": 0.782,
    "args": {
      "External id": 532157,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3211904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233413.574, "dur": 0.757,
    "args": {
      "External id": 532158,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3342976"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233417.938, "dur": 0.513,
    "args": {
      "External id": 532159,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3474048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233419.757, "dur": 0.771,
    "args": {
      "External id": 532160,"Record function id": 0, "Concrete Inputs": ["", "[131072]", "[1]", "3605120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233422.126, "dur": 0.680,
    "args": {
      "External id": 532161,"Record function id": 0, "Concrete Inputs": ["", "[128]", "[1]", "3736192"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233425.880, "dur": 0.592,
    "args": {
      "External id": 532162,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "3736320"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233428.004, "dur": 0.672,
    "args": {
      "External id": 532163,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4096768"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622233429.962, "dur": 2.435,
    "args": {
      "External id": 532164,"Record function id": 0, "Concrete Inputs": ["", "[360448]", "[1]", "4457216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[1605888], [], [], []], "Ev Idx": 6323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622233449.276, "dur": 21.695,
    "args": {
      "External id": 532165,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], [[128], [131072], [131072], [131072], [131072], [128], [360448], [360448], [360448]], []], "Ev Idx": 6324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 4183438, "tid": 4183438,
    "ts": 679622233532.994, "dur": 406.611,
    "args": {
      "External id": 532166,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[12847104], [1605888], [], [], []], "Ev Idx": 6325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622233559.707, "dur": 374.527,
    "args": {
      "External id": 532167,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 2, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 12847104, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[1605888], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6326, "In msg nelems": 1605888
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 4183438, "tid": 4183438,
    "ts": 679622233570.923, "dur": 356.705,
    "args": {
      "External id": 532168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[1605888]], "Ev Idx": 6327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622233962.721, "dur": 2.285,
    "args": {
      "External id": 532169,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6328, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.19)", "pid": 4183438, "tid": 4183438,
    "ts": 679622234059.504, "dur": 7683.203,
    "args": {
      "External id": 532170,"Record function id": 0, "Ev Idx": 6329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234177.448, "dur": 7.175,
    "args": {
      "External id": 532171,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[12847104], []], "Ev Idx": 6330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234197.577, "dur": 1.765,
    "args": {
      "External id": 532172,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234201.052, "dur": 1.228,
    "args": {
      "External id": 532173,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234204.044, "dur": 0.889,
    "args": {
      "External id": 532174,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234206.473, "dur": 1.240,
    "args": {
      "External id": 532175,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234209.074, "dur": 1.226,
    "args": {
      "External id": 532176,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1048576], []], "Ev Idx": 6335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234212.390, "dur": 0.829,
    "args": {
      "External id": 532177,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[1024], []], "Ev Idx": 6336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234215.035, "dur": 1.974,
    "args": {
      "External id": 532178,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234218.914, "dur": 0.740,
    "args": {
      "External id": 532179,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622234221.365, "dur": 0.865,
    "args": {
      "External id": 532180,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2883584], []], "Ev Idx": 6339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622234243.240, "dur": 7459.665,
    "args": {
      "External id": 532181,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622234259.519, "dur": 7436.056,
    "args": {
      "External id": 532182,"Record function id": 0, "Concrete Inputs": ["", "[128, 131072, 131072, 131072, 131072, 128, 360448, 360448, 360448]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[1605888, 1], [], [], [[128, 1], [131072, 1], [131072, 1], [131072, 1], [131072, 1], [128, 1], [360448, 1], [360448, 1], [360448, 1]]], "Input Dims": [[8, 1605888], [], [], [[8, 128], [8, 131072], [8, 131072], [8, 131072], [8, 131072], [8, 128], [8, 360448], [8, 360448], [8, 360448]]], "Ev Idx": 6341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622234274.027, "dur": 14.504,
    "args": {
      "External id": 532183,"Record function id": 0, "Concrete Inputs": ["[823]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622234291.199, "dur": 7349.866,
    "args": {
      "External id": 532184,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], [], []], "Ev Idx": 6343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622234293.780, "dur": 7346.728,
    "args": {
      "External id": 532185,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[823], [], [], [], [], [], []], "Ev Idx": 6344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622234299.500, "dur": 5.841,
    "args": {
      "External id": 532186,"Record function id": 0, "Concrete Inputs": ["[823]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622234307.163, "dur": 7330.635,
    "args": {
      "External id": 532187,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[823], [823], []], "Ev Idx": 6346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622241883.709, "dur": 30.272,
    "args": {
      "External id": 532188,"Sequence number": 6319455, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024]], "Ev Idx": 6347
    }
  },
  {
    "ph": "s", "id": 173, "pid": 4183438, "tid": 4183438, "ts": 679622241883.709,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622241901.254, "dur": 7.800,
    "args": {
      "External id": 532189,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1]], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024]], "Ev Idx": 6348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622241904.303, "dur": 4.495,
    "args": {
      "External id": 532190,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622241978.559, "dur": 88.364,
    "args": {
      "External id": 532191,"Record function id": 0, "Ev Idx": 6350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 4183438, "tid": 4183438,
    "ts": 679622242068.728, "dur": 1081.158,
    "args": {
      "External id": 532192,"Record function id": 0, "Ev Idx": 6351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622242123.172, "dur": 1011.501,
    "args": {
      "External id": 532193,"Sequence number": 6319456, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1], [1024, 1], [1024, 1], [1024, 1], [32, 1], [32, 1], [1024, 1], [1], [1024, 1], [1024, 1], [2816, 1]], "Input Dims": [[16, 4096, 1024], [1024], [1024, 1024], [1024, 1024], [1024, 1024], [8192, 32], [8192, 32], [1024, 1024], [1024], [2816, 1024], [2816, 1024], [1024, 2816]], "Ev Idx": 6352
    }
  },
  {
    "ph": "s", "id": 172, "pid": 4183438, "tid": 4183438, "ts": 679622242123.172,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622242192.054, "dur": 44.258,
    "args": {
      "External id": 532194,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622242251.293, "dur": 94.014,
    "args": {
      "External id": 532195,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622242354.998, "dur": 34.484,
    "args": {
      "External id": 532196,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622242398.762, "dur": 28.830,
    "args": {
      "External id": 532197,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622242451.602, "dur": 24.530,
    "args": {
      "External id": 532198,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622242493.579, "dur": 14.007,
    "args": {
      "External id": 532199,"kernel_hash": "cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "16", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/v3/cv3g2wsdzeyx5boypyj27xaitsqinaivjy4szifyb3uqvdwkozw5.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622242524.987, "dur": 164.060,
    "args": {
      "External id": 532200,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622242575.946, "dur": 11.804,
    "args": {
      "External id": 532201,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622242581.214, "dur": 5.521,
    "args": {
      "External id": 532202,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622242590.592, "dur": 4.500,
    "args": {
      "External id": 532203,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622242596.583, "dur": 1.637,
    "args": {
      "External id": 532204,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622242600.648, "dur": 3.045,
    "args": {
      "External id": 532205,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622242703.532, "dur": 56.429,
    "args": {
      "External id": 532206,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [1024, 1]], "Input Dims": [[65536, 1024], [1024, 1024], [65536, 1024]], "Ev Idx": 6365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 4183438, "tid": 4183438,
    "ts": 679622242793.857, "dur": 27.179,
    "args": {
      "External id": 532207,"kernel_hash": "cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/tt/cttakkz2bnieukw67monti4duhgj3fsbcaibwt2mjgnnrdznnjm2.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1024, 1], [1024, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536, 1024], [1024], [65536, 1024], [65536, 1024], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622242829.563, "dur": 38.453,
    "args": {
      "External id": 532208,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622242876.986, "dur": 32.300,
    "args": {
      "External id": 532209,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024], [2816, 1]], "Input Dims": [[65536, 1024], [1024, 2816], [65536, 2816]], "Ev Idx": 6368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622242932.356, "dur": 25.689,
    "args": {
      "External id": 532210,"kernel_hash": "clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp", "grid": "grid(184549376,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "184549376"], "kernel_file": "/tmp/torchinductor_cvm/lz/clzix6hvzjx55eyxbe3uw2ngyqaq3mu6uwbzhejcgoxnzlj4utbp.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2816, 1], [2816, 1], [11534336, 2816, 1], []], "Input Dims": [[65536, 2816], [65536, 2816], [16, 4096, 2816], []], "Ev Idx": 6369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622242964.523, "dur": 43.189,
    "args": {
      "External id": 532211,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816], [1024, 1]], "Input Dims": [[65536, 2816], [2816, 1024], [65536, 1024]], "Ev Idx": 6370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 4183438, "tid": 4183438,
    "ts": 679622243026.180, "dur": 19.937,
    "args": {
      "External id": 532212,"kernel_hash": "cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds", "grid": "grid(67108864,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "67108864"], "kernel_file": "/tmp/torchinductor_cvm/im/cimycee4ybs4iko3ux3z53wborga7phhsjwuofkutdacu75wrlds.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [65536, 1024], []], "Ev Idx": 6371
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.19)", "pid": 4183438, "tid": 4183438,
    "ts": 679622243216.440, "dur": 33.051,
    "args": {
      "External id": 532213,"Record function id": 0, "Ev Idx": 6372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622243330.757, "dur": 262.255,
    "args": {
      "External id": 532214,"Sequence number": 6319457, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [], [], [], [], []], "Ev Idx": 6373
    }
  },
  {
    "ph": "s", "id": 171, "pid": 4183438, "tid": 4183438, "ts": 679622243330.757,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243358.406, "dur": 8.016,
    "args": {
      "External id": 532215,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243360.579, "dur": 5.605,
    "args": {
      "External id": 532216,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622243376.923, "dur": 12.453,
    "args": {
      "External id": 532217,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622243380.583, "dur": 8.134,
    "args": {
      "External id": 532218,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622243396.788, "dur": 4.674,
    "args": {
      "External id": 532219,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243573.864, "dur": 6.930,
    "args": {
      "External id": 532220,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243577.983, "dur": 2.568,
    "args": {
      "External id": 532221,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622243621.693, "dur": 174.794,
    "args": {
      "External id": 532222,"Sequence number": 6319458, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622243624.739, "dur": 14.957,
    "args": {
      "External id": 532223,"Sequence number": 6319458, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6382
    }
  },
  {
    "ph": "s", "id": 170, "pid": 4183438, "tid": 4183438, "ts": 679622243624.739,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622243631.262, "dur": 6.788,
    "args": {
      "External id": 532224,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622243635.561, "dur": 2.023,
    "args": {
      "External id": 532225,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622243642.585, "dur": 153.494,
    "args": {
      "External id": 532226,"Sequence number": 6319459, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243645.732, "dur": 5.234,
    "args": {
      "External id": 532227,"Sequence number": 6319459, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243647.157, "dur": 3.654,
    "args": {
      "External id": 532228,"Sequence number": 6319459, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6387
    }
  },
  {
    "ph": "s", "id": 169, "pid": 4183438, "tid": 4183438, "ts": 679622243647.157,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622243690.153, "dur": 94.958,
    "args": {
      "External id": 532229,"Sequence number": 6319460, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6388
    }
  },
  {
    "ph": "s", "id": 168, "pid": 4183438, "tid": 4183438, "ts": 679622243690.153,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243789.028, "dur": 5.740,
    "args": {
      "External id": 532230,"Sequence number": 6319461, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6389
    }
  },
  {
    "ph": "s", "id": 167, "pid": 4183438, "tid": 4183438, "ts": 679622243789.028,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622243808.515, "dur": 67.991,
    "args": {
      "External id": 532231,"Sequence number": 6319462, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622243809.779, "dur": 7.793,
    "args": {
      "External id": 532232,"Sequence number": 6319462, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6391
    }
  },
  {
    "ph": "s", "id": 166, "pid": 4183438, "tid": 4183438, "ts": 679622243809.779,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622243812.145, "dur": 3.964,
    "args": {
      "External id": 532233,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622243814.576, "dur": 1.168,
    "args": {
      "External id": 532234,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622243818.692, "dur": 57.560,
    "args": {
      "External id": 532235,"Sequence number": 6319463, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243820.672, "dur": 6.686,
    "args": {
      "External id": 532236,"Sequence number": 6319463, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243822.182, "dur": 5.014,
    "args": {
      "External id": 532237,"Sequence number": 6319463, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6396
    }
  },
  {
    "ph": "s", "id": 165, "pid": 4183438, "tid": 4183438, "ts": 679622243822.182,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622243828.478, "dur": 40.979,
    "args": {
      "External id": 532238,"Sequence number": 6319464, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6397
    }
  },
  {
    "ph": "s", "id": 164, "pid": 4183438, "tid": 4183438, "ts": 679622243828.478,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243871.359, "dur": 4.425,
    "args": {
      "External id": 532239,"Sequence number": 6319465, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6398
    }
  },
  {
    "ph": "s", "id": 163, "pid": 4183438, "tid": 4183438, "ts": 679622243871.359,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622243884.114, "dur": 60.500,
    "args": {
      "External id": 532240,"Sequence number": 6319466, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622243885.137, "dur": 6.769,
    "args": {
      "External id": 532241,"Sequence number": 6319466, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6400
    }
  },
  {
    "ph": "s", "id": 162, "pid": 4183438, "tid": 4183438, "ts": 679622243885.137,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622243887.869, "dur": 3.027,
    "args": {
      "External id": 532242,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622243889.503, "dur": 1.146,
    "args": {
      "External id": 532243,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622243892.701, "dur": 51.722,
    "args": {
      "External id": 532244,"Sequence number": 6319467, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243894.081, "dur": 5.110,
    "args": {
      "External id": 532245,"Sequence number": 6319467, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243895.346, "dur": 3.686,
    "args": {
      "External id": 532246,"Sequence number": 6319467, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6405
    }
  },
  {
    "ph": "s", "id": 161, "pid": 4183438, "tid": 4183438, "ts": 679622243895.346,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622243900.122, "dur": 37.895,
    "args": {
      "External id": 532247,"Sequence number": 6319468, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6406
    }
  },
  {
    "ph": "s", "id": 160, "pid": 4183438, "tid": 4183438, "ts": 679622243900.122,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243939.743, "dur": 4.299,
    "args": {
      "External id": 532248,"Sequence number": 6319469, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6407
    }
  },
  {
    "ph": "s", "id": 159, "pid": 4183438, "tid": 4183438, "ts": 679622243939.743,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243963.749, "dur": 4.813,
    "args": {
      "External id": 532249,"Sequence number": 6319470, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243965.499, "dur": 2.918,
    "args": {
      "External id": 532250,"Sequence number": 6319470, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6409
    }
  },
  {
    "ph": "s", "id": 158, "pid": 4183438, "tid": 4183438, "ts": 679622243965.499,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243979.076, "dur": 5.405,
    "args": {
      "External id": 532251,"Sequence number": 6319471, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243980.229, "dur": 4.108,
    "args": {
      "External id": 532252,"Sequence number": 6319471, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6411
    }
  },
  {
    "ph": "s", "id": 157, "pid": 4183438, "tid": 4183438, "ts": 679622243980.229,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622243988.991, "dur": 3.939,
    "args": {
      "External id": 532253,"Sequence number": 6319472, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622243990.566, "dur": 2.057,
    "args": {
      "External id": 532254,"Sequence number": 6319472, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6413
    }
  },
  {
    "ph": "s", "id": 156, "pid": 4183438, "tid": 4183438, "ts": 679622243990.566,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622244028.482, "dur": 202.269,
    "args": {
      "External id": 532255,"Sequence number": 6319473, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6414
    }
  },
  {
    "ph": "s", "id": 155, "pid": 4183438, "tid": 4183438, "ts": 679622244028.482,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622244049.620, "dur": 8.932,
    "args": {
      "External id": 532256,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622244052.567, "dur": 5.532,
    "args": {
      "External id": 532257,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622244250.107, "dur": 112.599,
    "args": {
      "External id": 532258,"Sequence number": 6319474, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6417
    }
  },
  {
    "ph": "s", "id": 154, "pid": 4183438, "tid": 4183438, "ts": 679622244250.107,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622244264.762, "dur": 10.216,
    "args": {
      "External id": 532259,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622244267.914, "dur": 6.640,
    "args": {
      "External id": 532260,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 4183438, "tid": 4183438,
    "ts": 679622244395.331, "dur": 181.869,
    "args": {
      "External id": 532261,"Sequence number": 6319475, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 6420
    }
  },
  {
    "ph": "s", "id": 153, "pid": 4183438, "tid": 4183438, "ts": 679622244395.331,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622244424.100, "dur": 123.449,
    "args": {
      "External id": 532262,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622244472.504, "dur": 6.980,
    "args": {
      "External id": 532263,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622244475.009, "dur": 4.062,
    "args": {
      "External id": 532264,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622244482.139, "dur": 4.145,
    "args": {
      "External id": 532265,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622244487.811, "dur": 1.444,
    "args": {
      "External id": 532266,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622244491.902, "dur": 4.914,
    "args": {
      "External id": 532267,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 4183438,
    "ts": 679622244561.511, "dur": 6.209,
    "args": {
      "External id": 532268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 6427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622244582.690, "dur": 6.544,
    "args": {
      "External id": 532269,"Sequence number": 6319476, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622244584.337, "dur": 4.752,
    "args": {
      "External id": 532270,"Sequence number": 6319476, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6429
    }
  },
  {
    "ph": "s", "id": 152, "pid": 4183438, "tid": 4183438, "ts": 679622244584.337,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622244601.540, "dur": 157.837,
    "args": {
      "External id": 532271,"Sequence number": 6319477, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622244603.183, "dur": 9.649,
    "args": {
      "External id": 532272,"Sequence number": 6319477, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6431
    }
  },
  {
    "ph": "s", "id": 151, "pid": 4183438, "tid": 4183438, "ts": 679622244603.183,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622244607.151, "dur": 4.530,
    "args": {
      "External id": 532273,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622244609.767, "dur": 1.583,
    "args": {
      "External id": 532274,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622244614.487, "dur": 144.496,
    "args": {
      "External id": 532275,"Sequence number": 6319478, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622244616.755, "dur": 4.900,
    "args": {
      "External id": 532276,"Sequence number": 6319478, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622244618.276, "dur": 3.236,
    "args": {
      "External id": 532277,"Sequence number": 6319478, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6436
    }
  },
  {
    "ph": "s", "id": 150, "pid": 4183438, "tid": 4183438, "ts": 679622244618.276,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622244622.944, "dur": 127.246,
    "args": {
      "External id": 532278,"Sequence number": 6319479, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6437
    }
  },
  {
    "ph": "s", "id": 149, "pid": 4183438, "tid": 4183438, "ts": 679622244622.944,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622244754.425, "dur": 3.774,
    "args": {
      "External id": 532279,"Sequence number": 6319480, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6438
    }
  },
  {
    "ph": "s", "id": 148, "pid": 4183438, "tid": 4183438, "ts": 679622244754.425,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622244800.694, "dur": 229.928,
    "args": {
      "External id": 532280,"Sequence number": 6319481, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [16, 4096, 1024], [], [], [], []], "Ev Idx": 6439
    }
  },
  {
    "ph": "s", "id": 147, "pid": 4183438, "tid": 4183438, "ts": 679622244800.694,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622244821.541, "dur": 3.723,
    "args": {
      "External id": 532281,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622244823.148, "dur": 1.959,
    "args": {
      "External id": 532282,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622244830.494, "dur": 3.700,
    "args": {
      "External id": 532283,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1024, 1]], "Input Dims": [[16, 4096, 1024], [65536, 1024]], "Ev Idx": 6442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622244831.867, "dur": 2.220,
    "args": {
      "External id": 532284,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622244832.916, "dur": 1.078,
    "args": {
      "External id": 532285,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622244842.592, "dur": 8.717,
    "args": {
      "External id": 532286,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622244845.207, "dur": 5.716,
    "args": {
      "External id": 532287,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622244858.465, "dur": 3.620,
    "args": {
      "External id": 532288,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622244865.688, "dur": 4.326,
    "args": {
      "External id": 532289,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245006.230, "dur": 6.154,
    "args": {
      "External id": 532290,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245007.474, "dur": 4.677,
    "args": {
      "External id": 532291,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245015.201, "dur": 2.617,
    "args": {
      "External id": 532292,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245016.536, "dur": 1.168,
    "args": {
      "External id": 532293,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622245051.195, "dur": 136.891,
    "args": {
      "External id": 532294,"Sequence number": 6319482, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622245052.422, "dur": 8.850,
    "args": {
      "External id": 532295,"Sequence number": 6319482, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6454
    }
  },
  {
    "ph": "s", "id": 146, "pid": 4183438, "tid": 4183438, "ts": 679622245052.422,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622245055.811, "dur": 4.364,
    "args": {
      "External id": 532296,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622245058.284, "dur": 1.596,
    "args": {
      "External id": 532297,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622245062.828, "dur": 124.990,
    "args": {
      "External id": 532298,"Sequence number": 6319483, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245064.936, "dur": 5.910,
    "args": {
      "External id": 532299,"Sequence number": 6319483, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245066.271, "dur": 4.452,
    "args": {
      "External id": 532300,"Sequence number": 6319483, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6459
    }
  },
  {
    "ph": "s", "id": 145, "pid": 4183438, "tid": 4183438, "ts": 679622245066.271,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622245071.758, "dur": 108.394,
    "args": {
      "External id": 532301,"Sequence number": 6319484, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6460
    }
  },
  {
    "ph": "s", "id": 144, "pid": 4183438, "tid": 4183438, "ts": 679622245071.758,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245183.545, "dur": 3.624,
    "args": {
      "External id": 532302,"Sequence number": 6319485, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6461
    }
  },
  {
    "ph": "s", "id": 143, "pid": 4183438, "tid": 4183438, "ts": 679622245183.545,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622245198.511, "dur": 70.985,
    "args": {
      "External id": 532303,"Sequence number": 6319486, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622245199.452, "dur": 8.083,
    "args": {
      "External id": 532304,"Sequence number": 6319486, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6463
    }
  },
  {
    "ph": "s", "id": 142, "pid": 4183438, "tid": 4183438, "ts": 679622245199.452,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622245201.218, "dur": 5.103,
    "args": {
      "External id": 532305,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622245205.196, "dur": 0.931,
    "args": {
      "External id": 532306,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622245208.671, "dur": 60.583,
    "args": {
      "External id": 532307,"Sequence number": 6319487, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245210.025, "dur": 5.644,
    "args": {
      "External id": 532308,"Sequence number": 6319487, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245211.466, "dur": 4.052,
    "args": {
      "External id": 532309,"Sequence number": 6319487, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6468
    }
  },
  {
    "ph": "s", "id": 141, "pid": 4183438, "tid": 4183438, "ts": 679622245211.466,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622245216.400, "dur": 46.626,
    "args": {
      "External id": 532310,"Sequence number": 6319488, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6469
    }
  },
  {
    "ph": "s", "id": 140, "pid": 4183438, "tid": 4183438, "ts": 679622245216.400,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245265.022, "dur": 3.867,
    "args": {
      "External id": 532311,"Sequence number": 6319489, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6470
    }
  },
  {
    "ph": "s", "id": 139, "pid": 4183438, "tid": 4183438, "ts": 679622245265.022,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622245292.997, "dur": 162.704,
    "args": {
      "External id": 532312,"Sequence number": 6319490, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6471
    }
  },
  {
    "ph": "s", "id": 138, "pid": 4183438, "tid": 4183438, "ts": 679622245292.997,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622245338.316, "dur": 4.525,
    "args": {
      "External id": 532313,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622245379.981, "dur": 60.066,
    "args": {
      "External id": 532314,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622245380.830, "dur": 5.337,
    "args": {
      "External id": 532315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 6474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622245382.162, "dur": 3.124,
    "args": {
      "External id": 532316,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 6475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622245384.064, "dur": 0.936,
    "args": {
      "External id": 532317,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 6476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622245387.264, "dur": 52.503,
    "args": {
      "External id": 532318,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[11534336, 2816, 1], [1, 2816]], "Input Dims": [[16, 4096, 2816], [2816, 1024]], "Ev Idx": 6477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245389.024, "dur": 3.193,
    "args": {
      "External id": 532319,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245390.401, "dur": 1.688,
    "args": {
      "External id": 532320,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622245393.314, "dur": 42.541,
    "args": {
      "External id": 532321,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 6480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245437.552, "dur": 1.455,
    "args": {
      "External id": 532322,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622245466.107, "dur": 28.823,
    "args": {
      "External id": 532323,"Sequence number": 6319491, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 6482
    }
  },
  {
    "ph": "s", "id": 137, "pid": 4183438, "tid": 4183438, "ts": 679622245466.107,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622245529.581, "dur": 242.949,
    "args": {
      "External id": 532324,"Sequence number": 6319492, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [], [], [], [], []], "Ev Idx": 6483
    }
  },
  {
    "ph": "s", "id": 136, "pid": 4183438, "tid": 4183438, "ts": 679622245529.581,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245549.082, "dur": 3.410,
    "args": {
      "External id": 532325,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245550.456, "dur": 1.768,
    "args": {
      "External id": 532326,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622245561.086, "dur": 7.523,
    "args": {
      "External id": 532327,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622245563.767, "dur": 4.408,
    "args": {
      "External id": 532328,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622245574.662, "dur": 3.534,
    "args": {
      "External id": 532329,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245752.269, "dur": 6.673,
    "args": {
      "External id": 532330,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245755.816, "dur": 2.829,
    "args": {
      "External id": 532331,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622245793.003, "dur": 91.194,
    "args": {
      "External id": 532332,"Sequence number": 6319493, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622245794.316, "dur": 8.463,
    "args": {
      "External id": 532333,"Sequence number": 6319493, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6492
    }
  },
  {
    "ph": "s", "id": 135, "pid": 4183438, "tid": 4183438, "ts": 679622245794.316,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622245797.103, "dur": 4.232,
    "args": {
      "External id": 532334,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622245799.529, "dur": 1.621,
    "args": {
      "External id": 532335,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622245803.896, "dur": 80.063,
    "args": {
      "External id": 532336,"Sequence number": 6319494, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245806.190, "dur": 4.553,
    "args": {
      "External id": 532337,"Sequence number": 6319494, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245807.382, "dur": 3.209,
    "args": {
      "External id": 532338,"Sequence number": 6319494, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6497
    }
  },
  {
    "ph": "s", "id": 134, "pid": 4183438, "tid": 4183438, "ts": 679622245807.382,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622245812.145, "dur": 66.099,
    "args": {
      "External id": 532339,"Sequence number": 6319495, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6498
    }
  },
  {
    "ph": "s", "id": 133, "pid": 4183438, "tid": 4183438, "ts": 679622245812.145,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245880.387, "dur": 2.927,
    "args": {
      "External id": 532340,"Sequence number": 6319496, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6499
    }
  },
  {
    "ph": "s", "id": 132, "pid": 4183438, "tid": 4183438, "ts": 679622245880.387,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622245892.054, "dur": 64.812,
    "args": {
      "External id": 532341,"Sequence number": 6319497, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622245892.935, "dur": 8.284,
    "args": {
      "External id": 532342,"Sequence number": 6319497, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6501
    }
  },
  {
    "ph": "s", "id": 131, "pid": 4183438, "tid": 4183438, "ts": 679622245892.935,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622245897.337, "dur": 2.856,
    "args": {
      "External id": 532343,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622245899.165, "dur": 0.869,
    "args": {
      "External id": 532344,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622245902.088, "dur": 54.524,
    "args": {
      "External id": 532345,"Sequence number": 6319498, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245903.235, "dur": 6.058,
    "args": {
      "External id": 532346,"Sequence number": 6319498, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245904.504, "dur": 4.630,
    "args": {
      "External id": 532347,"Sequence number": 6319498, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6506
    }
  },
  {
    "ph": "s", "id": 130, "pid": 4183438, "tid": 4183438, "ts": 679622245904.504,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622245910.331, "dur": 39.790,
    "args": {
      "External id": 532348,"Sequence number": 6319499, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6507
    }
  },
  {
    "ph": "s", "id": 129, "pid": 4183438, "tid": 4183438, "ts": 679622245910.331,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245951.899, "dur": 4.330,
    "args": {
      "External id": 532349,"Sequence number": 6319500, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6508
    }
  },
  {
    "ph": "s", "id": 128, "pid": 4183438, "tid": 4183438, "ts": 679622245951.899,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622245963.664, "dur": 59.248,
    "args": {
      "External id": 532350,"Sequence number": 6319501, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622245964.503, "dur": 6.366,
    "args": {
      "External id": 532351,"Sequence number": 6319501, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6510
    }
  },
  {
    "ph": "s", "id": 127, "pid": 4183438, "tid": 4183438, "ts": 679622245964.503,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622245966.579, "dur": 3.092,
    "args": {
      "External id": 532352,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622245968.599, "dur": 0.940,
    "args": {
      "External id": 532353,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622245972.073, "dur": 50.644,
    "args": {
      "External id": 532354,"Sequence number": 6319502, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622245973.088, "dur": 4.546,
    "args": {
      "External id": 532355,"Sequence number": 6319502, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622245974.053, "dur": 3.425,
    "args": {
      "External id": 532356,"Sequence number": 6319502, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6515
    }
  },
  {
    "ph": "s", "id": 126, "pid": 4183438, "tid": 4183438, "ts": 679622245974.053,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622245978.568, "dur": 38.714,
    "args": {
      "External id": 532357,"Sequence number": 6319503, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6516
    }
  },
  {
    "ph": "s", "id": 125, "pid": 4183438, "tid": 4183438, "ts": 679622245978.568,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246018.805, "dur": 3.386,
    "args": {
      "External id": 532358,"Sequence number": 6319504, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6517
    }
  },
  {
    "ph": "s", "id": 124, "pid": 4183438, "tid": 4183438, "ts": 679622246018.805,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246038.705, "dur": 4.387,
    "args": {
      "External id": 532359,"Sequence number": 6319505, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246039.985, "dur": 2.972,
    "args": {
      "External id": 532360,"Sequence number": 6319505, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6519
    }
  },
  {
    "ph": "s", "id": 123, "pid": 4183438, "tid": 4183438, "ts": 679622246039.985,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246049.397, "dur": 5.211,
    "args": {
      "External id": 532361,"Sequence number": 6319506, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246050.658, "dur": 3.830,
    "args": {
      "External id": 532362,"Sequence number": 6319506, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6521
    }
  },
  {
    "ph": "s", "id": 122, "pid": 4183438, "tid": 4183438, "ts": 679622246050.658,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246058.904, "dur": 5.274,
    "args": {
      "External id": 532363,"Sequence number": 6319507, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246060.038, "dur": 3.992,
    "args": {
      "External id": 532364,"Sequence number": 6319507, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6523
    }
  },
  {
    "ph": "s", "id": 121, "pid": 4183438, "tid": 4183438, "ts": 679622246060.038,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622246093.520, "dur": 175.909,
    "args": {
      "External id": 532365,"Sequence number": 6319508, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6524
    }
  },
  {
    "ph": "s", "id": 120, "pid": 4183438, "tid": 4183438, "ts": 679622246093.520,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622246130.256, "dur": 9.332,
    "args": {
      "External id": 532366,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622246133.148, "dur": 5.834,
    "args": {
      "External id": 532367,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622246283.090, "dur": 95.037,
    "args": {
      "External id": 532368,"Sequence number": 6319509, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6527
    }
  },
  {
    "ph": "s", "id": 119, "pid": 4183438, "tid": 4183438, "ts": 679622246283.090,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622246295.657, "dur": 6.588,
    "args": {
      "External id": 532369,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622246297.989, "dur": 3.833,
    "args": {
      "External id": 532370,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 4183438, "tid": 4183438,
    "ts": 679622246405.192, "dur": 176.901,
    "args": {
      "External id": 532371,"Sequence number": 6319510, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 6530
    }
  },
  {
    "ph": "s", "id": 118, "pid": 4183438, "tid": 4183438, "ts": 679622246405.192,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622246430.715, "dur": 126.437,
    "args": {
      "External id": 532372,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622246478.355, "dur": 8.138,
    "args": {
      "External id": 532373,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622246481.080, "dur": 4.790,
    "args": {
      "External id": 532374,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622246489.313, "dur": 4.212,
    "args": {
      "External id": 532375,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622246495.124, "dur": 1.780,
    "args": {
      "External id": 532376,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622246501.854, "dur": 3.436,
    "args": {
      "External id": 532377,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 4183438,
    "ts": 679622246567.850, "dur": 5.112,
    "args": {
      "External id": 532378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 6537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246587.524, "dur": 6.535,
    "args": {
      "External id": 532379,"Sequence number": 6319511, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246589.379, "dur": 4.528,
    "args": {
      "External id": 532380,"Sequence number": 6319511, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6539
    }
  },
  {
    "ph": "s", "id": 117, "pid": 4183438, "tid": 4183438, "ts": 679622246589.379,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622246604.967, "dur": 151.524,
    "args": {
      "External id": 532381,"Sequence number": 6319512, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622246606.328, "dur": 9.001,
    "args": {
      "External id": 532382,"Sequence number": 6319512, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6541
    }
  },
  {
    "ph": "s", "id": 116, "pid": 4183438, "tid": 4183438, "ts": 679622246606.328,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622246609.224, "dur": 5.094,
    "args": {
      "External id": 532383,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622246612.482, "dur": 1.528,
    "args": {
      "External id": 532384,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622246616.795, "dur": 139.321,
    "args": {
      "External id": 532385,"Sequence number": 6319513, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246619.249, "dur": 3.546,
    "args": {
      "External id": 532386,"Sequence number": 6319513, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246620.429, "dur": 2.229,
    "args": {
      "External id": 532387,"Sequence number": 6319513, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6546
    }
  },
  {
    "ph": "s", "id": 115, "pid": 4183438, "tid": 4183438, "ts": 679622246620.429,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622246623.780, "dur": 123.173,
    "args": {
      "External id": 532388,"Sequence number": 6319514, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6547
    }
  },
  {
    "ph": "s", "id": 114, "pid": 4183438, "tid": 4183438, "ts": 679622246623.780,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246751.340, "dur": 3.973,
    "args": {
      "External id": 532389,"Sequence number": 6319515, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6548
    }
  },
  {
    "ph": "s", "id": 113, "pid": 4183438, "tid": 4183438, "ts": 679622246751.340,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622246793.989, "dur": 226.145,
    "args": {
      "External id": 532390,"Sequence number": 6319516, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [16, 4096, 1024], [], [], [], []], "Ev Idx": 6549
    }
  },
  {
    "ph": "s", "id": 112, "pid": 4183438, "tid": 4183438, "ts": 679622246793.989,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246815.770, "dur": 3.437,
    "args": {
      "External id": 532391,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246817.118, "dur": 1.941,
    "args": {
      "External id": 532392,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622246823.541, "dur": 3.935,
    "args": {
      "External id": 532393,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1024, 1]], "Input Dims": [[16, 4096, 1024], [65536, 1024]], "Ev Idx": 6552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246824.909, "dur": 2.460,
    "args": {
      "External id": 532394,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246826.332, "dur": 0.932,
    "args": {
      "External id": 532395,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622246835.725, "dur": 8.132,
    "args": {
      "External id": 532396,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622246838.239, "dur": 5.310,
    "args": {
      "External id": 532397,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622246850.818, "dur": 3.492,
    "args": {
      "External id": 532398,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622246857.943, "dur": 3.036,
    "args": {
      "External id": 532399,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622246994.830, "dur": 5.939,
    "args": {
      "External id": 532400,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622246998.573, "dur": 1.963,
    "args": {
      "External id": 532401,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247003.640, "dur": 2.704,
    "args": {
      "External id": 532402,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247004.927, "dur": 1.302,
    "args": {
      "External id": 532403,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622247037.804, "dur": 140.572,
    "args": {
      "External id": 532404,"Sequence number": 6319517, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622247039.052, "dur": 9.293,
    "args": {
      "External id": 532405,"Sequence number": 6319517, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6564
    }
  },
  {
    "ph": "s", "id": 111, "pid": 4183438, "tid": 4183438, "ts": 679622247039.052,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622247042.249, "dur": 4.682,
    "args": {
      "External id": 532406,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622247045.072, "dur": 1.605,
    "args": {
      "External id": 532407,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622247049.360, "dur": 128.693,
    "args": {
      "External id": 532408,"Sequence number": 6319518, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247051.332, "dur": 6.595,
    "args": {
      "External id": 532409,"Sequence number": 6319518, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247053.090, "dur": 4.657,
    "args": {
      "External id": 532410,"Sequence number": 6319518, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6569
    }
  },
  {
    "ph": "s", "id": 110, "pid": 4183438, "tid": 4183438, "ts": 679622247053.090,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622247058.619, "dur": 88.327,
    "args": {
      "External id": 532411,"Sequence number": 6319519, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6570
    }
  },
  {
    "ph": "s", "id": 109, "pid": 4183438, "tid": 4183438, "ts": 679622247058.619,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247166.038, "dur": 11.196,
    "args": {
      "External id": 532412,"Sequence number": 6319520, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6571
    }
  },
  {
    "ph": "s", "id": 108, "pid": 4183438, "tid": 4183438, "ts": 679622247166.038,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622247188.726, "dur": 72.589,
    "args": {
      "External id": 532413,"Sequence number": 6319521, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622247189.992, "dur": 6.662,
    "args": {
      "External id": 532414,"Sequence number": 6319521, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6573
    }
  },
  {
    "ph": "s", "id": 107, "pid": 4183438, "tid": 4183438, "ts": 679622247189.992,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622247191.931, "dur": 3.189,
    "args": {
      "External id": 532415,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622247193.887, "dur": 1.056,
    "args": {
      "External id": 532416,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622247197.609, "dur": 63.496,
    "args": {
      "External id": 532417,"Sequence number": 6319522, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247199.212, "dur": 5.250,
    "args": {
      "External id": 532418,"Sequence number": 6319522, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247200.469, "dur": 3.867,
    "args": {
      "External id": 532419,"Sequence number": 6319522, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6578
    }
  },
  {
    "ph": "s", "id": 106, "pid": 4183438, "tid": 4183438, "ts": 679622247200.469,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622247205.168, "dur": 51.673,
    "args": {
      "External id": 532420,"Sequence number": 6319523, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6579
    }
  },
  {
    "ph": "s", "id": 105, "pid": 4183438, "tid": 4183438, "ts": 679622247205.168,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247258.501, "dur": 2.228,
    "args": {
      "External id": 532421,"Sequence number": 6319524, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6580
    }
  },
  {
    "ph": "s", "id": 104, "pid": 4183438, "tid": 4183438, "ts": 679622247258.501,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622247286.775, "dur": 148.614,
    "args": {
      "External id": 532422,"Sequence number": 6319525, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6581
    }
  },
  {
    "ph": "s", "id": 103, "pid": 4183438, "tid": 4183438, "ts": 679622247286.775,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622247326.751, "dur": 4.536,
    "args": {
      "External id": 532423,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622247363.859, "dur": 58.067,
    "args": {
      "External id": 532424,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622247364.666, "dur": 6.815,
    "args": {
      "External id": 532425,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 6584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622247365.778, "dur": 4.907,
    "args": {
      "External id": 532426,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 6585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622247369.731, "dur": 0.775,
    "args": {
      "External id": 532427,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 6586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622247372.478, "dur": 49.166,
    "args": {
      "External id": 532428,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[11534336, 2816, 1], [1, 2816]], "Input Dims": [[16, 4096, 2816], [2816, 1024]], "Ev Idx": 6587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247374.203, "dur": 2.818,
    "args": {
      "External id": 532429,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247375.685, "dur": 1.204,
    "args": {
      "External id": 532430,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622247377.831, "dur": 40.242,
    "args": {
      "External id": 532431,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 6590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247419.896, "dur": 1.128,
    "args": {
      "External id": 532432,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622247444.308, "dur": 22.738,
    "args": {
      "External id": 532433,"Sequence number": 6319526, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 6592
    }
  },
  {
    "ph": "s", "id": 102, "pid": 4183438, "tid": 4183438, "ts": 679622247444.308,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622247502.203, "dur": 246.766,
    "args": {
      "External id": 532434,"Sequence number": 6319527, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [], [], [], [], []], "Ev Idx": 6593
    }
  },
  {
    "ph": "s", "id": 101, "pid": 4183438, "tid": 4183438, "ts": 679622247502.203,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247520.636, "dur": 3.725,
    "args": {
      "External id": 532435,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247522.022, "dur": 2.107,
    "args": {
      "External id": 532436,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622247531.924, "dur": 7.982,
    "args": {
      "External id": 532437,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622247534.921, "dur": 4.581,
    "args": {
      "External id": 532438,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622247545.923, "dur": 3.129,
    "args": {
      "External id": 532439,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247728.911, "dur": 4.494,
    "args": {
      "External id": 532440,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247730.142, "dur": 2.936,
    "args": {
      "External id": 532441,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622247768.914, "dur": 94.692,
    "args": {
      "External id": 532442,"Sequence number": 6319528, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622247770.282, "dur": 8.193,
    "args": {
      "External id": 532443,"Sequence number": 6319528, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6602
    }
  },
  {
    "ph": "s", "id": 100, "pid": 4183438, "tid": 4183438, "ts": 679622247770.282,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622247773.093, "dur": 4.009,
    "args": {
      "External id": 532444,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622247775.218, "dur": 1.627,
    "args": {
      "External id": 532445,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622247779.881, "dur": 83.398,
    "args": {
      "External id": 532446,"Sequence number": 6319529, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247781.722, "dur": 4.099,
    "args": {
      "External id": 532447,"Sequence number": 6319529, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247782.771, "dur": 2.889,
    "args": {
      "External id": 532448,"Sequence number": 6319529, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6607
    }
  },
  {
    "ph": "s", "id": 99, "pid": 4183438, "tid": 4183438, "ts": 679622247782.771,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622247786.764, "dur": 66.516,
    "args": {
      "External id": 532449,"Sequence number": 6319530, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6608
    }
  },
  {
    "ph": "s", "id": 98, "pid": 4183438, "tid": 4183438, "ts": 679622247786.764,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247855.583, "dur": 6.950,
    "args": {
      "External id": 532450,"Sequence number": 6319531, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6609
    }
  },
  {
    "ph": "s", "id": 97, "pid": 4183438, "tid": 4183438, "ts": 679622247855.583,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622247871.146, "dur": 59.624,
    "args": {
      "External id": 532451,"Sequence number": 6319532, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622247871.905, "dur": 6.268,
    "args": {
      "External id": 532452,"Sequence number": 6319532, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6611
    }
  },
  {
    "ph": "s", "id": 96, "pid": 4183438, "tid": 4183438, "ts": 679622247871.905,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622247874.103, "dur": 2.972,
    "args": {
      "External id": 532453,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622247875.965, "dur": 0.949,
    "args": {
      "External id": 532454,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622247879.611, "dur": 50.978,
    "args": {
      "External id": 532455,"Sequence number": 6319533, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247880.849, "dur": 3.997,
    "args": {
      "External id": 532456,"Sequence number": 6319533, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247882.426, "dur": 2.279,
    "args": {
      "External id": 532457,"Sequence number": 6319533, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6616
    }
  },
  {
    "ph": "s", "id": 95, "pid": 4183438, "tid": 4183438, "ts": 679622247882.426,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622247885.585, "dur": 40.572,
    "args": {
      "External id": 532458,"Sequence number": 6319534, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6617
    }
  },
  {
    "ph": "s", "id": 94, "pid": 4183438, "tid": 4183438, "ts": 679622247885.585,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247928.188, "dur": 2.066,
    "args": {
      "External id": 532459,"Sequence number": 6319535, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6618
    }
  },
  {
    "ph": "s", "id": 93, "pid": 4183438, "tid": 4183438, "ts": 679622247928.188,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622247937.769, "dur": 59.248,
    "args": {
      "External id": 532460,"Sequence number": 6319536, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622247938.733, "dur": 6.780,
    "args": {
      "External id": 532461,"Sequence number": 6319536, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6620
    }
  },
  {
    "ph": "s", "id": 92, "pid": 4183438, "tid": 4183438, "ts": 679622247938.733,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622247941.182, "dur": 3.196,
    "args": {
      "External id": 532462,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622247943.359, "dur": 0.870,
    "args": {
      "External id": 532463,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622247946.366, "dur": 50.463,
    "args": {
      "External id": 532464,"Sequence number": 6319537, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622247947.879, "dur": 5.392,
    "args": {
      "External id": 532465,"Sequence number": 6319537, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247949.437, "dur": 3.686,
    "args": {
      "External id": 532466,"Sequence number": 6319537, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6625
    }
  },
  {
    "ph": "s", "id": 91, "pid": 4183438, "tid": 4183438, "ts": 679622247949.437,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622247954.158, "dur": 37.743,
    "args": {
      "External id": 532467,"Sequence number": 6319538, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6626
    }
  },
  {
    "ph": "s", "id": 90, "pid": 4183438, "tid": 4183438, "ts": 679622247954.158,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622247993.817, "dur": 2.730,
    "args": {
      "External id": 532468,"Sequence number": 6319539, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6627
    }
  },
  {
    "ph": "s", "id": 89, "pid": 4183438, "tid": 4183438, "ts": 679622247993.817,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622248012.376, "dur": 3.873,
    "args": {
      "External id": 532469,"Sequence number": 6319540, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248013.427, "dur": 2.681,
    "args": {
      "External id": 532470,"Sequence number": 6319540, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6629
    }
  },
  {
    "ph": "s", "id": 88, "pid": 4183438, "tid": 4183438, "ts": 679622248013.427,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622248022.819, "dur": 5.700,
    "args": {
      "External id": 532471,"Sequence number": 6319541, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248024.156, "dur": 4.193,
    "args": {
      "External id": 532472,"Sequence number": 6319541, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6631
    }
  },
  {
    "ph": "s", "id": 87, "pid": 4183438, "tid": 4183438, "ts": 679622248024.156,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622248033.293, "dur": 3.703,
    "args": {
      "External id": 532473,"Sequence number": 6319542, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248034.493, "dur": 2.156,
    "args": {
      "External id": 532474,"Sequence number": 6319542, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6633
    }
  },
  {
    "ph": "s", "id": 86, "pid": 4183438, "tid": 4183438, "ts": 679622248034.493,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622248067.222, "dur": 176.157,
    "args": {
      "External id": 532475,"Sequence number": 6319543, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6634
    }
  },
  {
    "ph": "s", "id": 85, "pid": 4183438, "tid": 4183438, "ts": 679622248067.222,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622248087.619, "dur": 8.445,
    "args": {
      "External id": 532476,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622248090.490, "dur": 4.998,
    "args": {
      "External id": 532477,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622248257.490, "dur": 115.218,
    "args": {
      "External id": 532478,"Sequence number": 6319544, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6637
    }
  },
  {
    "ph": "s", "id": 84, "pid": 4183438, "tid": 4183438, "ts": 679622248257.490,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622248272.836, "dur": 8.269,
    "args": {
      "External id": 532479,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622248275.283, "dur": 5.445,
    "args": {
      "External id": 532480,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 4183438, "tid": 4183438,
    "ts": 679622248442.160, "dur": 180.352,
    "args": {
      "External id": 532481,"Sequence number": 6319545, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 6640
    }
  },
  {
    "ph": "s", "id": 83, "pid": 4183438, "tid": 4183438, "ts": 679622248442.160,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622248469.335, "dur": 127.724,
    "args": {
      "External id": 532482,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622248517.099, "dur": 9.071,
    "args": {
      "External id": 532483,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622248519.132, "dur": 6.364,
    "args": {
      "External id": 532484,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622248531.503, "dur": 3.969,
    "args": {
      "External id": 532485,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622248537.040, "dur": 1.483,
    "args": {
      "External id": 532486,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622248541.103, "dur": 3.612,
    "args": {
      "External id": 532487,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 4183438,
    "ts": 679622248608.179, "dur": 5.116,
    "args": {
      "External id": 532488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 6647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622248627.712, "dur": 5.650,
    "args": {
      "External id": 532489,"Sequence number": 6319546, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248629.052, "dur": 4.165,
    "args": {
      "External id": 532490,"Sequence number": 6319546, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6649
    }
  },
  {
    "ph": "s", "id": 82, "pid": 4183438, "tid": 4183438, "ts": 679622248629.052,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622248645.764, "dur": 166.484,
    "args": {
      "External id": 532491,"Sequence number": 6319547, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622248646.834, "dur": 51.339,
    "args": {
      "External id": 532492,"Sequence number": 6319547, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6651
    }
  },
  {
    "ph": "s", "id": 81, "pid": 4183438, "tid": 4183438, "ts": 679622248646.834,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622248651.747, "dur": 44.562,
    "args": {
      "External id": 532493,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622248693.418, "dur": 2.225,
    "args": {
      "External id": 532494,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622248700.206, "dur": 111.593,
    "args": {
      "External id": 532495,"Sequence number": 6319548, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622248702.777, "dur": 4.706,
    "args": {
      "External id": 532496,"Sequence number": 6319548, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248703.858, "dur": 3.486,
    "args": {
      "External id": 532497,"Sequence number": 6319548, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6656
    }
  },
  {
    "ph": "s", "id": 80, "pid": 4183438, "tid": 4183438, "ts": 679622248703.858,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622248708.304, "dur": 93.509,
    "args": {
      "External id": 532498,"Sequence number": 6319549, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6657
    }
  },
  {
    "ph": "s", "id": 79, "pid": 4183438, "tid": 4183438, "ts": 679622248708.304,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248804.783, "dur": 6.106,
    "args": {
      "External id": 532499,"Sequence number": 6319550, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6658
    }
  },
  {
    "ph": "s", "id": 78, "pid": 4183438, "tid": 4183438, "ts": 679622248804.783,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622248852.591, "dur": 220.286,
    "args": {
      "External id": 532500,"Sequence number": 6319551, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [16, 4096, 1024], [], [], [], []], "Ev Idx": 6659
    }
  },
  {
    "ph": "s", "id": 77, "pid": 4183438, "tid": 4183438, "ts": 679622248852.591,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622248872.622, "dur": 3.197,
    "args": {
      "External id": 532501,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248873.854, "dur": 1.729,
    "args": {
      "External id": 532502,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622248879.976, "dur": 3.625,
    "args": {
      "External id": 532503,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1024, 1]], "Input Dims": [[16, 4096, 1024], [65536, 1024]], "Ev Idx": 6662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622248881.355, "dur": 2.134,
    "args": {
      "External id": 532504,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622248882.513, "dur": 0.868,
    "args": {
      "External id": 532505,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622248891.912, "dur": 7.676,
    "args": {
      "External id": 532506,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622248894.593, "dur": 4.665,
    "args": {
      "External id": 532507,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622248906.398, "dur": 3.455,
    "args": {
      "External id": 532508,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622248913.286, "dur": 3.412,
    "args": {
      "External id": 532509,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249048.657, "dur": 3.736,
    "args": {
      "External id": 532510,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249050.118, "dur": 2.023,
    "args": {
      "External id": 532511,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249055.243, "dur": 2.834,
    "args": {
      "External id": 532512,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249056.763, "dur": 1.193,
    "args": {
      "External id": 532513,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622249091.681, "dur": 119.311,
    "args": {
      "External id": 532514,"Sequence number": 6319552, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622249093.125, "dur": 25.480,
    "args": {
      "External id": 532515,"Sequence number": 6319552, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6674
    }
  },
  {
    "ph": "s", "id": 76, "pid": 4183438, "tid": 4183438, "ts": 679622249093.125,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622249095.310, "dur": 21.610,
    "args": {
      "External id": 532516,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622249114.712, "dur": 1.650,
    "args": {
      "External id": 532517,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622249120.031, "dur": 90.556,
    "args": {
      "External id": 532518,"Sequence number": 6319553, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249122.299, "dur": 4.428,
    "args": {
      "External id": 532519,"Sequence number": 6319553, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249123.694, "dur": 2.891,
    "args": {
      "External id": 532520,"Sequence number": 6319553, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6679
    }
  },
  {
    "ph": "s", "id": 75, "pid": 4183438, "tid": 4183438, "ts": 679622249123.694,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622249127.898, "dur": 74.164,
    "args": {
      "External id": 532521,"Sequence number": 6319554, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6680
    }
  },
  {
    "ph": "s", "id": 74, "pid": 4183438, "tid": 4183438, "ts": 679622249127.898,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249204.321, "dur": 5.671,
    "args": {
      "External id": 532522,"Sequence number": 6319555, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6681
    }
  },
  {
    "ph": "s", "id": 73, "pid": 4183438, "tid": 4183438, "ts": 679622249204.321,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622249221.353, "dur": 68.863,
    "args": {
      "External id": 532523,"Sequence number": 6319556, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622249222.470, "dur": 6.899,
    "args": {
      "External id": 532524,"Sequence number": 6319556, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6683
    }
  },
  {
    "ph": "s", "id": 72, "pid": 4183438, "tid": 4183438, "ts": 679622249222.470,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622249224.649, "dur": 3.377,
    "args": {
      "External id": 532525,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622249226.797, "dur": 1.065,
    "args": {
      "External id": 532526,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622249230.558, "dur": 59.280,
    "args": {
      "External id": 532527,"Sequence number": 6319557, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249232.262, "dur": 7.993,
    "args": {
      "External id": 532528,"Sequence number": 6319557, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249233.868, "dur": 6.098,
    "args": {
      "External id": 532529,"Sequence number": 6319557, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6688
    }
  },
  {
    "ph": "s", "id": 71, "pid": 4183438, "tid": 4183438, "ts": 679622249233.868,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622249240.946, "dur": 42.742,
    "args": {
      "External id": 532530,"Sequence number": 6319558, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6689
    }
  },
  {
    "ph": "s", "id": 70, "pid": 4183438, "tid": 4183438, "ts": 679622249240.946,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249284.999, "dur": 4.541,
    "args": {
      "External id": 532531,"Sequence number": 6319559, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6690
    }
  },
  {
    "ph": "s", "id": 69, "pid": 4183438, "tid": 4183438, "ts": 679622249284.999,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622249312.531, "dur": 144.364,
    "args": {
      "External id": 532532,"Sequence number": 6319560, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6691
    }
  },
  {
    "ph": "s", "id": 68, "pid": 4183438, "tid": 4183438, "ts": 679622249312.531,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622249350.412, "dur": 4.041,
    "args": {
      "External id": 532533,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622249386.050, "dur": 57.289,
    "args": {
      "External id": 532534,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622249386.889, "dur": 4.690,
    "args": {
      "External id": 532535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 6694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622249387.915, "dur": 2.966,
    "args": {
      "External id": 532536,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 6695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622249389.738, "dur": 0.958,
    "args": {
      "External id": 532537,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 6696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622249392.267, "dur": 50.783,
    "args": {
      "External id": 532538,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[11534336, 2816, 1], [1, 2816]], "Input Dims": [[16, 4096, 2816], [2816, 1024]], "Ev Idx": 6697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249394.304, "dur": 2.754,
    "args": {
      "External id": 532539,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249395.484, "dur": 1.447,
    "args": {
      "External id": 532540,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622249397.870, "dur": 41.603,
    "args": {
      "External id": 532541,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 6700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249441.400, "dur": 1.016,
    "args": {
      "External id": 532542,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622249465.006, "dur": 26.049,
    "args": {
      "External id": 532543,"Sequence number": 6319561, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 6702
    }
  },
  {
    "ph": "s", "id": 67, "pid": 4183438, "tid": 4183438, "ts": 679622249465.006,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622249526.193, "dur": 224.859,
    "args": {
      "External id": 532544,"Sequence number": 6319562, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [], [], [], [], []], "Ev Idx": 6703
    }
  },
  {
    "ph": "s", "id": 66, "pid": 4183438, "tid": 4183438, "ts": 679622249526.193,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249543.633, "dur": 3.729,
    "args": {
      "External id": 532545,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249545.192, "dur": 1.976,
    "args": {
      "External id": 532546,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622249554.252, "dur": 7.889,
    "args": {
      "External id": 532547,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622249557.172, "dur": 4.461,
    "args": {
      "External id": 532548,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622249567.761, "dur": 3.893,
    "args": {
      "External id": 532549,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249733.060, "dur": 4.467,
    "args": {
      "External id": 532550,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249734.182, "dur": 3.094,
    "args": {
      "External id": 532551,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622249770.999, "dur": 93.790,
    "args": {
      "External id": 532552,"Sequence number": 6319563, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622249772.336, "dur": 7.767,
    "args": {
      "External id": 532553,"Sequence number": 6319563, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6712
    }
  },
  {
    "ph": "s", "id": 65, "pid": 4183438, "tid": 4183438, "ts": 679622249772.336,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622249774.692, "dur": 4.180,
    "args": {
      "External id": 532554,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622249777.012, "dur": 1.639,
    "args": {
      "External id": 532555,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622249781.040, "dur": 83.485,
    "args": {
      "External id": 532556,"Sequence number": 6319564, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249782.979, "dur": 4.625,
    "args": {
      "External id": 532557,"Sequence number": 6319564, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249784.389, "dur": 3.082,
    "args": {
      "External id": 532558,"Sequence number": 6319564, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6717
    }
  },
  {
    "ph": "s", "id": 64, "pid": 4183438, "tid": 4183438, "ts": 679622249784.389,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622249788.466, "dur": 69.965,
    "args": {
      "External id": 532559,"Sequence number": 6319565, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6718
    }
  },
  {
    "ph": "s", "id": 63, "pid": 4183438, "tid": 4183438, "ts": 679622249788.466,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249860.908, "dur": 3.037,
    "args": {
      "External id": 532560,"Sequence number": 6319566, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6719
    }
  },
  {
    "ph": "s", "id": 62, "pid": 4183438, "tid": 4183438, "ts": 679622249860.908,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622249872.429, "dur": 63.562,
    "args": {
      "External id": 532561,"Sequence number": 6319567, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622249873.438, "dur": 5.405,
    "args": {
      "External id": 532562,"Sequence number": 6319567, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6721
    }
  },
  {
    "ph": "s", "id": 61, "pid": 4183438, "tid": 4183438, "ts": 679622249873.438,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622249875.047, "dur": 2.809,
    "args": {
      "External id": 532563,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622249876.752, "dur": 0.972,
    "args": {
      "External id": 532564,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622249879.779, "dur": 56.023,
    "args": {
      "External id": 532565,"Sequence number": 6319568, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249881.265, "dur": 3.972,
    "args": {
      "External id": 532566,"Sequence number": 6319568, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249882.489, "dur": 2.617,
    "args": {
      "External id": 532567,"Sequence number": 6319568, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6726
    }
  },
  {
    "ph": "s", "id": 60, "pid": 4183438, "tid": 4183438, "ts": 679622249882.489,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622249886.001, "dur": 43.166,
    "args": {
      "External id": 532568,"Sequence number": 6319569, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6727
    }
  },
  {
    "ph": "s", "id": 59, "pid": 4183438, "tid": 4183438, "ts": 679622249886.001,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249930.944, "dur": 4.543,
    "args": {
      "External id": 532569,"Sequence number": 6319570, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6728
    }
  },
  {
    "ph": "s", "id": 58, "pid": 4183438, "tid": 4183438, "ts": 679622249930.944,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622249942.474, "dur": 67.358,
    "args": {
      "External id": 532570,"Sequence number": 6319571, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622249943.631, "dur": 7.220,
    "args": {
      "External id": 532571,"Sequence number": 6319571, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6730
    }
  },
  {
    "ph": "s", "id": 57, "pid": 4183438, "tid": 4183438, "ts": 679622249943.631,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622249945.317, "dur": 4.525,
    "args": {
      "External id": 532572,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622249947.071, "dur": 2.540,
    "args": {
      "External id": 532573,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622249951.716, "dur": 57.930,
    "args": {
      "External id": 532574,"Sequence number": 6319572, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622249952.978, "dur": 4.740,
    "args": {
      "External id": 532575,"Sequence number": 6319572, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622249954.272, "dur": 3.298,
    "args": {
      "External id": 532576,"Sequence number": 6319572, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6735
    }
  },
  {
    "ph": "s", "id": 56, "pid": 4183438, "tid": 4183438, "ts": 679622249954.272,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622249958.456, "dur": 45.601,
    "args": {
      "External id": 532577,"Sequence number": 6319573, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6736
    }
  },
  {
    "ph": "s", "id": 55, "pid": 4183438, "tid": 4183438, "ts": 679622249958.456,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250005.364, "dur": 3.907,
    "args": {
      "External id": 532578,"Sequence number": 6319574, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6737
    }
  },
  {
    "ph": "s", "id": 54, "pid": 4183438, "tid": 4183438, "ts": 679622250005.364,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250025.410, "dur": 4.514,
    "args": {
      "External id": 532579,"Sequence number": 6319575, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250026.261, "dur": 3.530,
    "args": {
      "External id": 532580,"Sequence number": 6319575, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6739
    }
  },
  {
    "ph": "s", "id": 53, "pid": 4183438, "tid": 4183438, "ts": 679622250026.261,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250036.746, "dur": 3.175,
    "args": {
      "External id": 532581,"Sequence number": 6319576, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250037.931, "dur": 1.867,
    "args": {
      "External id": 532582,"Sequence number": 6319576, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6741
    }
  },
  {
    "ph": "s", "id": 52, "pid": 4183438, "tid": 4183438, "ts": 679622250037.931,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250044.135, "dur": 3.376,
    "args": {
      "External id": 532583,"Sequence number": 6319577, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250045.440, "dur": 1.946,
    "args": {
      "External id": 532584,"Sequence number": 6319577, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 16, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6743
    }
  },
  {
    "ph": "s", "id": 51, "pid": 4183438, "tid": 4183438, "ts": 679622250045.440,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622250077.635, "dur": 170.106,
    "args": {
      "External id": 532585,"Sequence number": 6319578, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6744
    }
  },
  {
    "ph": "s", "id": 50, "pid": 4183438, "tid": 4183438, "ts": 679622250077.635,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622250095.445, "dur": 27.168,
    "args": {
      "External id": 532586,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622250114.378, "dur": 7.614,
    "args": {
      "External id": 532587,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622250268.890, "dur": 102.139,
    "args": {
      "External id": 532588,"Sequence number": 6319579, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 6747
    }
  },
  {
    "ph": "s", "id": 49, "pid": 4183438, "tid": 4183438, "ts": 679622250268.890,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622250284.401, "dur": 7.307,
    "args": {
      "External id": 532589,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622250286.638, "dur": 4.726,
    "args": {
      "External id": 532590,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 4183438, "tid": 4183438,
    "ts": 679622250398.098, "dur": 173.245,
    "args": {
      "External id": 532591,"Sequence number": 6319580, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 6750
    }
  },
  {
    "ph": "s", "id": 48, "pid": 4183438, "tid": 4183438, "ts": 679622250398.098,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622250425.283, "dur": 121.631,
    "args": {
      "External id": 532592,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [4194304, 1024, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [16, 4096, 16, 64], [16, 4096, 16, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622250473.731, "dur": 6.273,
    "args": {
      "External id": 532593,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[4194304, 1024, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 16, 64], [], [], [], [], []], "Ev Idx": 6752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622250475.835, "dur": 3.780,
    "args": {
      "External id": 532594,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 16, 64]", "[4194304, 1024, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622250482.869, "dur": 3.463,
    "args": {
      "External id": 532595,"Record function id": 0, "Concrete Inputs": ["[16, 16, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622250487.633, "dur": 1.740,
    "args": {
      "External id": 532596,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622250491.835, "dur": 5.512,
    "args": {
      "External id": 532597,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 4183438, "tid": 4183438,
    "ts": 679622250557.879, "dur": 4.639,
    "args": {
      "External id": 532598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[4194304, 1024, 64, 1]], "Input Dims": [[16, 4096, 16, 64]], "Ev Idx": 6757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250576.025, "dur": 5.593,
    "args": {
      "External id": 532599,"Sequence number": 6319581, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250577.636, "dur": 3.813,
    "args": {
      "External id": 532600,"Sequence number": 6319581, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 64, 1], []], "Input Dims": [[16, 4096, 16, 64], []], "Ev Idx": 6759
    }
  },
  {
    "ph": "s", "id": 47, "pid": 4183438, "tid": 4183438, "ts": 679622250577.636,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622250592.085, "dur": 145.548,
    "args": {
      "External id": 532601,"Sequence number": 6319582, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [1024, 1024], []], "Ev Idx": 6760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622250593.580, "dur": 7.736,
    "args": {
      "External id": 532602,"Sequence number": 6319582, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[1024, 1024]], "Ev Idx": 6761
    }
  },
  {
    "ph": "s", "id": 46, "pid": 4183438, "tid": 4183438, "ts": 679622250593.580,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622250596.309, "dur": 3.957,
    "args": {
      "External id": 532603,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[1024, 1024], [], []], "Ev Idx": 6762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622250598.753, "dur": 1.307,
    "args": {
      "External id": 532604,"Record function id": 0, "Concrete Inputs": ["", "[1024, 1024]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[1024, 1024], [], [], []], "Ev Idx": 6763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622250602.413, "dur": 134.822,
    "args": {
      "External id": 532605,"Sequence number": 6319583, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 1024]], "Ev Idx": 6764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250604.577, "dur": 3.624,
    "args": {
      "External id": 532606,"Sequence number": 6319583, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250605.457, "dur": 2.596,
    "args": {
      "External id": 532607,"Sequence number": 6319583, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6766
    }
  },
  {
    "ph": "s", "id": 45, "pid": 4183438, "tid": 4183438, "ts": 679622250605.457,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622250609.304, "dur": 118.785,
    "args": {
      "External id": 532608,"Sequence number": 6319584, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 1024]], "Ev Idx": 6767
    }
  },
  {
    "ph": "s", "id": 44, "pid": 4183438, "tid": 4183438, "ts": 679622250609.304,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250732.100, "dur": 4.486,
    "args": {
      "External id": 532609,"Sequence number": 6319585, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6768
    }
  },
  {
    "ph": "s", "id": 43, "pid": 4183438, "tid": 4183438, "ts": 679622250732.100,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622250774.610, "dur": 209.612,
    "args": {
      "External id": 532610,"Sequence number": 6319586, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4194304, 1024, 1], [1], [], [4194304, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [1024], [], [16, 4096, 1024], [], [], [], []], "Ev Idx": 6769
    }
  },
  {
    "ph": "s", "id": 42, "pid": 4183438, "tid": 4183438, "ts": 679622250774.610,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250792.511, "dur": 3.585,
    "args": {
      "External id": 532611,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250793.884, "dur": 2.044,
    "args": {
      "External id": 532612,"Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622250799.941, "dur": 4.157,
    "args": {
      "External id": 532613,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1024, 1]], "Input Dims": [[16, 4096, 1024], [65536, 1024]], "Ev Idx": 6772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250801.817, "dur": 2.180,
    "args": {
      "External id": 532614,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250802.603, "dur": 1.304,
    "args": {
      "External id": 532615,"Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622250811.886, "dur": 7.845,
    "args": {
      "External id": 532616,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622250814.388, "dur": 4.975,
    "args": {
      "External id": 532617,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622250826.181, "dur": 3.534,
    "args": {
      "External id": 532618,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622250833.304, "dur": 2.826,
    "args": {
      "External id": 532619,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250960.565, "dur": 6.010,
    "args": {
      "External id": 532620,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250961.672, "dur": 4.643,
    "args": {
      "External id": 532621,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622250968.971, "dur": 2.281,
    "args": {
      "External id": 532622,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622250970.122, "dur": 1.024,
    "args": {
      "External id": 532623,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622251001.838, "dur": 115.169,
    "args": {
      "External id": 532624,"Sequence number": 6319587, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622251003.226, "dur": 7.748,
    "args": {
      "External id": 532625,"Sequence number": 6319587, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6784
    }
  },
  {
    "ph": "s", "id": 41, "pid": 4183438, "tid": 4183438, "ts": 679622251003.226,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622251005.852, "dur": 3.955,
    "args": {
      "External id": 532626,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622251008.350, "dur": 1.170,
    "args": {
      "External id": 532627,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622251012.429, "dur": 104.314,
    "args": {
      "External id": 532628,"Sequence number": 6319588, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622251014.196, "dur": 5.908,
    "args": {
      "External id": 532629,"Sequence number": 6319588, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622251015.653, "dur": 4.298,
    "args": {
      "External id": 532630,"Sequence number": 6319588, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6789
    }
  },
  {
    "ph": "s", "id": 40, "pid": 4183438, "tid": 4183438, "ts": 679622251015.653,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622251021.314, "dur": 72.424,
    "args": {
      "External id": 532631,"Sequence number": 6319589, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6790
    }
  },
  {
    "ph": "s", "id": 39, "pid": 4183438, "tid": 4183438, "ts": 679622251021.314,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622251095.934, "dur": 19.755,
    "args": {
      "External id": 532632,"Sequence number": 6319590, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6791
    }
  },
  {
    "ph": "s", "id": 38, "pid": 4183438, "tid": 4183438, "ts": 679622251095.934,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622251127.760, "dur": 84.188,
    "args": {
      "External id": 532633,"Sequence number": 6319591, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[4194304, 1024, 1], [1024, 1], []], "Input Dims": [[16, 4096, 1024], [2816, 1024], []], "Ev Idx": 6792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622251129.054, "dur": 11.917,
    "args": {
      "External id": 532634,"Sequence number": 6319591, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[2816, 1024]], "Ev Idx": 6793
    }
  },
  {
    "ph": "s", "id": 37, "pid": 4183438, "tid": 4183438, "ts": 679622251129.054,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622251134.018, "dur": 5.456,
    "args": {
      "External id": 532635,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[2816, 1024], [], []], "Ev Idx": 6794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622251138.260, "dur": 1.060,
    "args": {
      "External id": 532636,"Record function id": 0, "Concrete Inputs": ["", "[1024, 2816]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[2816, 1024], [], [], []], "Ev Idx": 6795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622251142.182, "dur": 69.561,
    "args": {
      "External id": 532637,"Sequence number": 6319592, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[4194304, 1024, 1], [1, 1024]], "Input Dims": [[16, 4096, 1024], [1024, 2816]], "Ev Idx": 6796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622251143.660, "dur": 5.589,
    "args": {
      "External id": 532638,"Sequence number": 6319592, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622251145.080, "dur": 3.837,
    "args": {
      "External id": 532639,"Sequence number": 6319592, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6798
    }
  },
  {
    "ph": "s", "id": 36, "pid": 4183438, "tid": 4183438, "ts": 679622251145.080,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622251150.028, "dur": 53.986,
    "args": {
      "External id": 532640,"Sequence number": 6319593, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[65536, 1024], [1024, 2816]], "Ev Idx": 6799
    }
  },
  {
    "ph": "s", "id": 35, "pid": 4183438, "tid": 4183438, "ts": 679622251150.028,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622251205.649, "dur": 5.749,
    "args": {
      "External id": 532641,"Sequence number": 6319594, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2816, 1], []], "Input Dims": [[65536, 2816], []], "Ev Idx": 6800
    }
  },
  {
    "ph": "s", "id": 34, "pid": 4183438, "tid": 4183438, "ts": 679622251205.649,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622251232.603, "dur": 148.882,
    "args": {
      "External id": 532642,"Sequence number": 6319595, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6801
    }
  },
  {
    "ph": "s", "id": 33, "pid": 4183438, "tid": 4183438, "ts": 679622251232.603,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622251272.153, "dur": 4.448,
    "args": {
      "External id": 532643,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2816]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622251307.912, "dur": 59.699,
    "args": {
      "External id": 532644,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[11534336, 2816, 1], [2816, 1], []], "Input Dims": [[16, 4096, 2816], [1024, 2816], []], "Ev Idx": 6803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622251308.721, "dur": 5.237,
    "args": {
      "External id": 532645,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2816, 1]], "Input Dims": [[1024, 2816]], "Ev Idx": 6804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622251310.067, "dur": 3.007,
    "args": {
      "External id": 532646,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2816, 1], [], []], "Input Dims": [[1024, 2816], [], []], "Ev Idx": 6805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622251311.902, "dur": 0.968,
    "args": {
      "External id": 532647,"Record function id": 0, "Concrete Inputs": ["", "[2816, 1024]", "[1, 2816]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2816, 1], [], [], []], "Input Dims": [[1024, 2816], [], [], []], "Ev Idx": 6806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622251314.675, "dur": 52.636,
    "args": {
      "External id": 532648,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[11534336, 2816, 1], [1, 2816]], "Input Dims": [[16, 4096, 2816], [2816, 1024]], "Ev Idx": 6807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622251316.418, "dur": 2.889,
    "args": {
      "External id": 532649,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622251317.690, "dur": 1.492,
    "args": {
      "External id": 532650,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2816]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 2816, 1], []], "Input Dims": [[16, 4096, 2816], []], "Ev Idx": 6809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622251320.193, "dur": 43.213,
    "args": {
      "External id": 532651,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2816, 1], [1, 2816]], "Input Dims": [[65536, 2816], [2816, 1024]], "Ev Idx": 6810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 4183438, "tid": 4183438,
    "ts": 679622251365.376, "dur": 1.294,
    "args": {
      "External id": 532652,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622251389.688, "dur": 25.748,
    "args": {
      "External id": 532653,"Sequence number": 6319596, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4194304, 1024, 1], [4194304, 1024, 1], []], "Input Dims": [[16, 4096, 1024], [16, 4096, 1024], []], "Ev Idx": 6812
    }
  },
  {
    "ph": "s", "id": 32, "pid": 4183438, "tid": 4183438, "ts": 679622251389.688,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 4183438, "tid": 4183438,
    "ts": 679622251434.884, "dur": 40.924,
    "args": {
      "External id": 532654,"Sequence number": 6319597, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-2"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[[4194304, 1024, 1], [4194304, 1024, 1], [4194304, 1024, 1], [4194304, 1024, 1]], []], "Input Dims": [[[16, 4096, 1024], [16, 4096, 1024], [16, 4096, 1024], [16, 4096, 1024]], []], "Ev Idx": 6813
    }
  },
  {
    "ph": "s", "id": 31, "pid": 4183438, "tid": 4183438, "ts": 679622251434.884,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 4183438, "tid": 4183438,
    "ts": 679622251442.867, "dur": 28.213,
    "args": {
      "External id": 532655,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[[4194304, 1024, 1], [4194304, 1024, 1], [4194304, 1024, 1], [4194304, 1024, 1]], []], "Input Dims": [[[16, 4096, 1024], [16, 4096, 1024], [16, 4096, 1024], [16, 4096, 1024]], []], "Ev Idx": 6814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622251472.680, "dur": 1.626,
    "args": {
      "External id": 532656,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 4096], []], "Ev Idx": 6815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 4183438, "tid": 4183438,
    "ts": 679622251509.211, "dur": 63.643,
    "args": {
      "External id": 532657,"Record function id": 0, "Ev Idx": 6816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 2/0", "pid": 4183438, "tid": 4183438,
    "ts": 679622251574.097, "dur": 234.128,
    "args": {
      "External id": 532658,"Record function id": 0, "Ev Idx": 6817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622251607.796, "dur": 190.839,
    "args": {
      "External id": 532659,"Sequence number": 6319598, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1], [16777216, 4096, 1024, 1]], "Input Dims": [[1024], [16, 4096, 4, 1024]], "Ev Idx": 6818
    }
  },
  {
    "ph": "s", "id": 30, "pid": 4183438, "tid": 4183438, "ts": 679622251607.796,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 4183438, "tid": 4183438,
    "ts": 679622251716.494, "dur": 40.184,
    "args": {
      "External id": 532660,"kernel_hash": "ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "1024", "1", "9.9999999999999995e-07", "True", "1024", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/cg/ccgi3dilqw7bx5zirazjzjmgxoha7dpoyovs4nylqgmksmgrgn6b.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[262144, 1024], [262144, 1024], [1024], [262144], [], [], [], [], [], [], [], [], []], "Ev Idx": 6819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622251889.608, "dur": 46.686,
    "args": {
      "External id": 532661,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False"], "Input type": ["ScalarList", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 6820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622251892.691, "dur": 5.843,
    "args": {
      "External id": 532662,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622251901.101, "dur": 34.870,
    "args": {
      "External id": 532663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 6822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622251904.568, "dur": 30.868,
    "args": {
      "External id": 532664,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 6823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622251941.009, "dur": 22.424,
    "args": {
      "External id": 532665,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False"], "Input type": ["ScalarList", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 6824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622251942.262, "dur": 3.342,
    "args": {
      "External id": 532666,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622251946.469, "dur": 16.677,
    "args": {
      "External id": 532667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 6826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622251947.153, "dur": 15.483,
    "args": {
      "External id": 532668,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 6827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622251966.739, "dur": 18.207,
    "args": {
      "External id": 532669,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False"], "Input type": ["ScalarList", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 6828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622251967.779, "dur": 2.646,
    "args": {
      "External id": 532670,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622251971.333, "dur": 13.353,
    "args": {
      "External id": 532671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 6830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622251972.357, "dur": 11.775,
    "args": {
      "External id": 532672,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 6831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622251994.652, "dur": 0.666,
    "args": {
      "External id": 532673,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], [], [], [], []], "Ev Idx": 6832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unfold", "pid": 4183438, "tid": 4183438,
    "ts": 679622252003.341, "dur": 9.802,
    "args": {
      "External id": 532674,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "5", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 6833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252009.418, "dur": 1.874,
    "args": {
      "External id": 532675,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 5]", "[8192, 1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 6834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252019.205, "dur": 7.133,
    "args": {
      "External id": 532676,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 5], [], [], [], []], "Ev Idx": 6835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252023.311, "dur": 1.282,
    "args": {
      "External id": 532677,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 5]", "[8192, 1, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 5], [], [], []], "Ev Idx": 6836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252027.588, "dur": 4.056,
    "args": {
      "External id": 532678,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 5], [], [], [], []], "Ev Idx": 6837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252029.745, "dur": 1.203,
    "args": {
      "External id": 532679,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 5]", "[8192, 1, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 5], [], [], []], "Ev Idx": 6838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252033.558, "dur": 3.476,
    "args": {
      "External id": 532680,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "1", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 5], [], [], [], []], "Ev Idx": 6839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252035.850, "dur": 0.586,
    "args": {
      "External id": 532681,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 5], [], [], []], "Ev Idx": 6840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252040.231, "dur": 3.143,
    "args": {
      "External id": 532682,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 4], [], [], [], []], "Ev Idx": 6841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252042.277, "dur": 0.534,
    "args": {
      "External id": 532683,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 4], [], [], []], "Ev Idx": 6842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252044.638, "dur": 3.574,
    "args": {
      "External id": 532684,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "4096", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 4], [], [], [], []], "Ev Idx": 6843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252046.631, "dur": 0.891,
    "args": {
      "External id": 532685,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 4], [], [], []], "Ev Idx": 6844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252049.319, "dur": 3.073,
    "args": {
      "External id": 532686,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4096, 4], [], [], [], []], "Ev Idx": 6845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252051.144, "dur": 0.751,
    "args": {
      "External id": 532687,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4096, 4], [], [], []], "Ev Idx": 6846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622252056.310, "dur": 6.013,
    "args": {
      "External id": 532688,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4096, 4], [], []], "Ev Idx": 6847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252060.864, "dur": 0.734,
    "args": {
      "External id": 532689,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4096, 4], [], [], []], "Ev Idx": 6848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252066.488, "dur": 3.508,
    "args": {
      "External id": 532690,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 6849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252068.562, "dur": 0.875,
    "args": {
      "External id": 532691,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 6850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622252073.324, "dur": 7.542,
    "args": {
      "External id": 532692,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 6851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252078.807, "dur": 0.999,
    "args": {
      "External id": 532693,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 6852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252082.057, "dur": 3.934,
    "args": {
      "External id": 532694,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 6853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252084.389, "dur": 1.065,
    "args": {
      "External id": 532695,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 6854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252088.635, "dur": 6.478,
    "args": {
      "External id": 532696,"Sequence number": 6319599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 6855
    }
  },
  {
    "ph": "s", "id": 29, "pid": 4183438, "tid": 4183438, "ts": 679622252088.635,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252092.436, "dur": 1.124,
    "args": {
      "External id": 532697,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 6856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252096.216, "dur": 23.224,
    "args": {
      "External id": 532698,"Sequence number": 6319600, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 6857
    }
  },
  {
    "ph": "s", "id": 28, "pid": 4183438, "tid": 4183438, "ts": 679622252096.216,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252116.872, "dur": 0.988,
    "args": {
      "External id": 532699,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 6858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622252120.996, "dur": 6.217,
    "args": {
      "External id": 532700,"Sequence number": 6319601, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 6859
    }
  },
  {
    "ph": "s", "id": 27, "pid": 4183438, "tid": 4183438, "ts": 679622252120.996,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252125.551, "dur": 0.816,
    "args": {
      "External id": 532701,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 6860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622252128.524, "dur": 5.002,
    "args": {
      "External id": 532702,"Sequence number": 6319602, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 6861
    }
  },
  {
    "ph": "s", "id": 26, "pid": 4183438, "tid": 4183438, "ts": 679622252128.524,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252131.609, "dur": 1.172,
    "args": {
      "External id": 532703,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 6862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622252139.017, "dur": 44.740,
    "args": {
      "External id": 532704,"Sequence number": 6319603, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 6863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622252142.171, "dur": 41.309,
    "args": {
      "External id": 532705,"Sequence number": 6319603, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 6864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622252144.768, "dur": 9.046,
    "args": {
      "External id": 532706,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 6865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622252147.453, "dur": 5.702,
    "args": {
      "External id": 532707,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252155.631, "dur": 27.077,
    "args": {
      "External id": 532708,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 6867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622252207.159, "dur": 4.701,
    "args": {
      "External id": 532709,"Sequence number": 6319603, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 6868
    }
  },
  {
    "ph": "s", "id": 25, "pid": 4183438, "tid": 4183438, "ts": 679622252207.159,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622252214.355, "dur": 1.305,
    "args": {
      "External id": 532710,"Sequence number": 6319604, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 6869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622252242.613, "dur": 21301.502,
    "args": {
      "External id": 532711,"Sequence number": 6319604, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1], [1], [1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536], [32000, 1024], [], [], [], [], []], "Ev Idx": 6870
    }
  },
  {
    "ph": "s", "id": 24, "pid": 4183438, "tid": 4183438, "ts": 679622252242.613,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622252256.558, "dur": 28.903,
    "args": {
      "External id": 532712,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622252257.432, "dur": 27.793,
    "args": {
      "External id": 532713,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622252258.946, "dur": 5.766,
    "args": {
      "External id": 532714,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[4096, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622252260.824, "dur": 3.520,
    "args": {
      "External id": 532715,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252265.538, "dur": 19.171,
    "args": {
      "External id": 532716,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [4096, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 6875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622252301.995, "dur": 36.233,
    "args": {
      "External id": 532717,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622252311.179, "dur": 7.297,
    "args": {
      "External id": 532718,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 6877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252314.139, "dur": 4.032,
    "args": {
      "External id": 532719,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252319.599, "dur": 18.416,
    "args": {
      "External id": 532720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 6879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252321.274, "dur": 16.398,
    "args": {
      "External id": 532721,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 6880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622252341.810, "dur": 23.507,
    "args": {
      "External id": 532722,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 6881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622252342.930, "dur": 5.529,
    "args": {
      "External id": 532723,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 6882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252344.804, "dur": 3.375,
    "args": {
      "External id": 532724,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252349.186, "dur": 15.915,
    "args": {
      "External id": 532725,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 6884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252349.823, "dur": 14.840,
    "args": {
      "External id": 532726,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[32000, 1024], []], "Ev Idx": 6885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622252369.633, "dur": 18.865,
    "args": {
      "External id": 532727,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 6886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622252371.035, "dur": 3.141,
    "args": {
      "External id": 532728,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252375.234, "dur": 12.985,
    "args": {
      "External id": 532729,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 6888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252376.063, "dur": 11.529,
    "args": {
      "External id": 532730,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 6889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 4183438,
    "ts": 679622252395.761, "dur": 27.529,
    "args": {
      "External id": 532731,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 6890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622252428.888, "dur": 63.612,
    "args": {
      "External id": 532732,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 6891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622252433.612, "dur": 58.391,
    "args": {
      "External id": 532733,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 6892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252440.788, "dur": 1.142,
    "args": {
      "External id": 532734,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 6893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622252444.005, "dur": 27.798,
    "args": {
      "External id": 532735,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 6894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622252448.393, "dur": 23.208,
    "args": {
      "External id": 532736,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 6895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622252451.321, "dur": 2.897,
    "args": {
      "External id": 532737,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622252454.871, "dur": 16.328,
    "args": {
      "External id": 532738,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 6897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622252498.536, "dur": 15128.145,
    "args": {
      "External id": 532739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622252500.642, "dur": 15124.755,
    "args": {
      "External id": 532740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622267638.673, "dur": 7.402,
    "args": {
      "External id": 532741,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 6900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622267643.016, "dur": 1.347,
    "args": {
      "External id": 532742,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 6901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622267673.735, "dur": 114.242,
    "args": {
      "External id": 532743,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 6902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622267675.916, "dur": 7.802,
    "args": {
      "External id": 532744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 6903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622267678.686, "dur": 4.128,
    "args": {
      "External id": 532745,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 6904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622267681.102, "dur": 1.303,
    "args": {
      "External id": 532746,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 6905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622267685.511, "dur": 101.488,
    "args": {
      "External id": 532747,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 6906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622267687.646, "dur": 98.355,
    "args": {
      "External id": 532748,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 6907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622267792.647, "dur": 4.706,
    "args": {
      "External id": 532749,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 6908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622267794.811, "dur": 1.121,
    "args": {
      "External id": 532750,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 6909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622267805.953, "dur": 3.896,
    "args": {
      "External id": 532751,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 6910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622267820.832, "dur": 8.198,
    "args": {
      "External id": 532752,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 6911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622267824.548, "dur": 4.240,
    "args": {
      "External id": 532753,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622267972.472, "dur": 276.030,
    "args": {
      "External id": 532754,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 6913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622267976.381, "dur": 4.433,
    "args": {
      "External id": 532755,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622267983.851, "dur": 264.046,
    "args": {
      "External id": 532756,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 6915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622267987.268, "dur": 0.723,
    "args": {
      "External id": 532757,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 6916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622267990.846, "dur": 32.097,
    "args": {
      "External id": 532758,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 6917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622268025.621, "dur": 7.104,
    "args": {
      "External id": 532759,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 6918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268031.344, "dur": 0.901,
    "args": {
      "External id": 532760,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 6919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622268034.498, "dur": 32.624,
    "args": {
      "External id": 532761,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 6920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622268035.938, "dur": 1.772,
    "args": {
      "External id": 532762,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622268040.268, "dur": 26.519,
    "args": {
      "External id": 532763,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 6922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268045.324, "dur": 4.208,
    "args": {
      "External id": 532764,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 6923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622268069.979, "dur": 26.301,
    "args": {
      "External id": 532765,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 6924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268113.194, "dur": 21.134,
    "args": {
      "External id": 532766,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 6925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622268138.738, "dur": 20.417,
    "args": {
      "External id": 532767,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 6926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268162.195, "dur": 19.349,
    "args": {
      "External id": 532768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 6927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622268185.130, "dur": 28.987,
    "args": {
      "External id": 532769,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 6928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268190.232, "dur": 2.585,
    "args": {
      "External id": 532770,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 6929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268195.832, "dur": 1.044,
    "args": {
      "External id": 532771,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 6930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268216.609, "dur": 15.344,
    "args": {
      "External id": 532772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 6931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268234.469, "dur": 12.266,
    "args": {
      "External id": 532773,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 6932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622268256.650, "dur": 2.622,
    "args": {
      "External id": 532774,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 6933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622268267.533, "dur": 5.324,
    "args": {
      "External id": 532775,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 6934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268270.659, "dur": 1.185,
    "args": {
      "External id": 532776,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 6935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622268358.902, "dur": 66.402,
    "args": {
      "External id": 532777,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 6936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622268431.234, "dur": 5.187,
    "args": {
      "External id": 532778,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 6937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268434.350, "dur": 0.682,
    "args": {
      "External id": 532779,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 6938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268438.737, "dur": 28.887,
    "args": {
      "External id": 532780,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 6939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622268474.038, "dur": 7.252,
    "args": {
      "External id": 532781,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 6940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622268476.123, "dur": 4.260,
    "args": {
      "External id": 532782,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 6941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268478.781, "dur": 1.386,
    "args": {
      "External id": 532783,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 6942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622268485.418, "dur": 55.024,
    "args": {
      "External id": 532784,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 6943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622268487.080, "dur": 52.782,
    "args": {
      "External id": 532785,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 6944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268545.359, "dur": 18.802,
    "args": {
      "External id": 532786,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 6945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622268570.061, "dur": 4.544,
    "args": {
      "External id": 532787,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 6946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268572.419, "dur": 1.192,
    "args": {
      "External id": 532788,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 6947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622268579.275, "dur": 59.615,
    "args": {
      "External id": 532789,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 6948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622268580.464, "dur": 11.361,
    "args": {
      "External id": 532790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 6949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622268581.663, "dur": 9.480,
    "args": {
      "External id": 532791,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 6950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268589.868, "dur": 1.132,
    "args": {
      "External id": 532792,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 6951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622268592.652, "dur": 45.733,
    "args": {
      "External id": 532793,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 6952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622268593.407, "dur": 44.390,
    "args": {
      "External id": 532794,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 6953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622268644.287, "dur": 4.307,
    "args": {
      "External id": 532795,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 6954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268646.559, "dur": 0.884,
    "args": {
      "External id": 532796,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 6955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622268693.323, "dur": 2.538,
    "args": {
      "External id": 532797,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 6956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622268706.277, "dur": 10.125,
    "args": {
      "External id": 532798,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 6957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622268709.217, "dur": 6.884,
    "args": {
      "External id": 532799,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622268821.808, "dur": 254.367,
    "args": {
      "External id": 532800,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 6959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622268824.249, "dur": 2.719,
    "args": {
      "External id": 532801,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622268828.735, "dur": 246.940,
    "args": {
      "External id": 532802,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 6961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622268831.125, "dur": 0.347,
    "args": {
      "External id": 532803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 6962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622268832.967, "dur": 28.775,
    "args": {
      "External id": 532804,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 6963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622268863.540, "dur": 4.175,
    "args": {
      "External id": 532805,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 6964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622268866.180, "dur": 1.140,
    "args": {
      "External id": 532806,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 6965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622268869.193, "dur": 34.474,
    "args": {
      "External id": 532807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 6966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622268870.548, "dur": 2.013,
    "args": {
      "External id": 532808,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622268873.867, "dur": 29.431,
    "args": {
      "External id": 532809,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 6968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268879.765, "dur": 2.995,
    "args": {
      "External id": 532810,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 6969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622268905.359, "dur": 38.833,
    "args": {
      "External id": 532811,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 6970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622268946.235, "dur": 28.052,
    "args": {
      "External id": 532812,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 6971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622268977.279, "dur": 24.432,
    "args": {
      "External id": 532813,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 6972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269003.241, "dur": 16.309,
    "args": {
      "External id": 532814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 6973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622269021.718, "dur": 23.778,
    "args": {
      "External id": 532815,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 6974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269024.024, "dur": 2.070,
    "args": {
      "External id": 532816,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 6975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269028.716, "dur": 0.914,
    "args": {
      "External id": 532817,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 6976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269047.826, "dur": 13.216,
    "args": {
      "External id": 532818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 6977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269062.406, "dur": 12.299,
    "args": {
      "External id": 532819,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 6978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622269083.082, "dur": 2.079,
    "args": {
      "External id": 532820,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 6979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622269094.621, "dur": 20.169,
    "args": {
      "External id": 532821,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 6980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269112.214, "dur": 0.995,
    "args": {
      "External id": 532822,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 6981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622269188.823, "dur": 53.956,
    "args": {
      "External id": 532823,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 6982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622269248.288, "dur": 6.552,
    "args": {
      "External id": 532824,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 6983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269251.747, "dur": 1.731,
    "args": {
      "External id": 532825,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 6984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269256.443, "dur": 26.690,
    "args": {
      "External id": 532826,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 6985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622269288.213, "dur": 19.565,
    "args": {
      "External id": 532827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 6986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622269295.776, "dur": 11.185,
    "args": {
      "External id": 532828,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 6987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269305.662, "dur": 1.121,
    "args": {
      "External id": 532829,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 6988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622269310.569, "dur": 45.020,
    "args": {
      "External id": 532830,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 6989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622269311.843, "dur": 42.965,
    "args": {
      "External id": 532831,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 6990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269359.669, "dur": 17.112,
    "args": {
      "External id": 532832,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 6991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622269382.193, "dur": 5.062,
    "args": {
      "External id": 532833,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 6992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269385.106, "dur": 1.291,
    "args": {
      "External id": 532834,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 6993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622269391.600, "dur": 54.582,
    "args": {
      "External id": 532835,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 6994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622269392.648, "dur": 7.044,
    "args": {
      "External id": 532836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 6995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622269393.767, "dur": 5.307,
    "args": {
      "External id": 532837,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 6996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269395.827, "dur": 2.928,
    "args": {
      "External id": 532838,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 6997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622269400.498, "dur": 45.189,
    "args": {
      "External id": 532839,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 6998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622269401.500, "dur": 43.406,
    "args": {
      "External id": 532840,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 6999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622269450.693, "dur": 4.541,
    "args": {
      "External id": 532841,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269453.224, "dur": 0.841,
    "args": {
      "External id": 532842,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622269461.325, "dur": 1.572,
    "args": {
      "External id": 532843,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622269471.049, "dur": 6.265,
    "args": {
      "External id": 532844,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622269473.441, "dur": 3.527,
    "args": {
      "External id": 532845,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622269562.463, "dur": 251.914,
    "args": {
      "External id": 532846,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622269564.960, "dur": 2.747,
    "args": {
      "External id": 532847,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622269569.408, "dur": 244.469,
    "args": {
      "External id": 532848,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622269570.553, "dur": 0.587,
    "args": {
      "External id": 532849,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622269575.051, "dur": 22.297,
    "args": {
      "External id": 532850,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622269598.904, "dur": 4.759,
    "args": {
      "External id": 532851,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269601.729, "dur": 1.514,
    "args": {
      "External id": 532852,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622269604.739, "dur": 25.009,
    "args": {
      "External id": 532853,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622269606.239, "dur": 3.410,
    "args": {
      "External id": 532854,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622269611.244, "dur": 18.253,
    "args": {
      "External id": 532855,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269614.044, "dur": 3.123,
    "args": {
      "External id": 532856,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622269631.113, "dur": 71.547,
    "args": {
      "External id": 532857,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269706.017, "dur": 19.035,
    "args": {
      "External id": 532858,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622269728.434, "dur": 15.850,
    "args": {
      "External id": 532859,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269745.892, "dur": 13.618,
    "args": {
      "External id": 532860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622269761.799, "dur": 23.336,
    "args": {
      "External id": 532861,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269764.045, "dur": 1.834,
    "args": {
      "External id": 532862,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269768.595, "dur": 0.900,
    "args": {
      "External id": 532863,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269787.183, "dur": 12.289,
    "args": {
      "External id": 532864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269801.125, "dur": 11.530,
    "args": {
      "External id": 532865,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622269822.045, "dur": 2.677,
    "args": {
      "External id": 532866,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622269834.595, "dur": 4.035,
    "args": {
      "External id": 532867,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269837.155, "dur": 0.561,
    "args": {
      "External id": 532868,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622269911.941, "dur": 49.388,
    "args": {
      "External id": 532869,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622269966.270, "dur": 5.363,
    "args": {
      "External id": 532870,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622269969.252, "dur": 1.307,
    "args": {
      "External id": 532871,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622269973.167, "dur": 23.985,
    "args": {
      "External id": 532872,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622270001.899, "dur": 8.353,
    "args": {
      "External id": 532873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622270003.886, "dur": 5.606,
    "args": {
      "External id": 532874,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270006.329, "dur": 2.898,
    "args": {
      "External id": 532875,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622270013.384, "dur": 44.773,
    "args": {
      "External id": 532876,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622270014.659, "dur": 42.735,
    "args": {
      "External id": 532877,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270062.183, "dur": 16.021,
    "args": {
      "External id": 532878,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622270083.177, "dur": 4.554,
    "args": {
      "External id": 532879,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270085.834, "dur": 0.861,
    "args": {
      "External id": 532880,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622270091.977, "dur": 69.583,
    "args": {
      "External id": 532881,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622270093.089, "dur": 18.894,
    "args": {
      "External id": 532882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622270094.094, "dur": 16.610,
    "args": {
      "External id": 532883,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270095.823, "dur": 1.001,
    "args": {
      "External id": 532884,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622270113.494, "dur": 47.585,
    "args": {
      "External id": 532885,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622270114.474, "dur": 46.019,
    "args": {
      "External id": 532886,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622270166.851, "dur": 4.474,
    "args": {
      "External id": 532887,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270169.120, "dur": 1.034,
    "args": {
      "External id": 532888,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622270177.661, "dur": 1.798,
    "args": {
      "External id": 532889,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270187.673, "dur": 7.483,
    "args": {
      "External id": 532890,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270190.578, "dur": 4.309,
    "args": {
      "External id": 532891,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622270280.991, "dur": 182.482,
    "args": {
      "External id": 532892,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270283.655, "dur": 2.651,
    "args": {
      "External id": 532893,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622270290.389, "dur": 172.511,
    "args": {
      "External id": 532894,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622270291.645, "dur": 0.564,
    "args": {
      "External id": 532895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622270293.764, "dur": 20.980,
    "args": {
      "External id": 532896,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622270316.375, "dur": 5.613,
    "args": {
      "External id": 532897,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270318.842, "dur": 2.877,
    "args": {
      "External id": 532898,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622270323.354, "dur": 22.283,
    "args": {
      "External id": 532899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270324.639, "dur": 1.433,
    "args": {
      "External id": 532900,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622270327.408, "dur": 17.889,
    "args": {
      "External id": 532901,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270330.851, "dur": 3.038,
    "args": {
      "External id": 532902,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622270347.310, "dur": 19.605,
    "args": {
      "External id": 532903,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270368.535, "dur": 12.410,
    "args": {
      "External id": 532904,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622270383.884, "dur": 13.870,
    "args": {
      "External id": 532905,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270399.469, "dur": 12.011,
    "args": {
      "External id": 532906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622270413.160, "dur": 23.041,
    "args": {
      "External id": 532907,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270415.137, "dur": 1.730,
    "args": {
      "External id": 532908,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270419.217, "dur": 2.500,
    "args": {
      "External id": 532909,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270437.998, "dur": 11.522,
    "args": {
      "External id": 532910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270450.628, "dur": 11.144,
    "args": {
      "External id": 532911,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622270469.684, "dur": 1.799,
    "args": {
      "External id": 532912,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622270480.306, "dur": 3.587,
    "args": {
      "External id": 532913,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270482.367, "dur": 0.612,
    "args": {
      "External id": 532914,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622270545.601, "dur": 41.944,
    "args": {
      "External id": 532915,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622270592.185, "dur": 4.982,
    "args": {
      "External id": 532916,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270595.095, "dur": 0.866,
    "args": {
      "External id": 532917,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270598.457, "dur": 23.488,
    "args": {
      "External id": 532918,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622270626.588, "dur": 6.206,
    "args": {
      "External id": 532919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622270628.303, "dur": 3.781,
    "args": {
      "External id": 532920,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270630.688, "dur": 1.202,
    "args": {
      "External id": 532921,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622270635.196, "dur": 95.457,
    "args": {
      "External id": 532922,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622270636.457, "dur": 92.905,
    "args": {
      "External id": 532923,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270736.341, "dur": 16.722,
    "args": {
      "External id": 532924,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622270759.468, "dur": 5.364,
    "args": {
      "External id": 532925,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270762.620, "dur": 0.998,
    "args": {
      "External id": 532926,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622270768.984, "dur": 50.661,
    "args": {
      "External id": 532927,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622270770.033, "dur": 5.010,
    "args": {
      "External id": 532928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622270771.507, "dur": 2.906,
    "args": {
      "External id": 532929,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270773.274, "dur": 0.939,
    "args": {
      "External id": 532930,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622270775.893, "dur": 43.372,
    "args": {
      "External id": 532931,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622270776.753, "dur": 42.019,
    "args": {
      "External id": 532932,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622270824.131, "dur": 3.628,
    "args": {
      "External id": 532933,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270826.025, "dur": 0.614,
    "args": {
      "External id": 532934,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622270833.690, "dur": 1.663,
    "args": {
      "External id": 532935,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270843.410, "dur": 6.850,
    "args": {
      "External id": 532936,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270845.995, "dur": 3.987,
    "args": {
      "External id": 532937,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622270939.996, "dur": 207.672,
    "args": {
      "External id": 532938,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270945.126, "dur": 1.735,
    "args": {
      "External id": 532939,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622270948.376, "dur": 198.581,
    "args": {
      "External id": 532940,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622270949.664, "dur": 0.498,
    "args": {
      "External id": 532941,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622270951.459, "dur": 23.177,
    "args": {
      "External id": 532942,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622270976.337, "dur": 5.803,
    "args": {
      "External id": 532943,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622270980.668, "dur": 1.239,
    "args": {
      "External id": 532944,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622270983.071, "dur": 23.904,
    "args": {
      "External id": 532945,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622270984.573, "dur": 1.850,
    "args": {
      "External id": 532946,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622270987.821, "dur": 18.817,
    "args": {
      "External id": 532947,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622270990.555, "dur": 2.849,
    "args": {
      "External id": 532948,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622271008.471, "dur": 20.544,
    "args": {
      "External id": 532949,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271030.592, "dur": 14.221,
    "args": {
      "External id": 532950,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622271047.099, "dur": 13.940,
    "args": {
      "External id": 532951,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271062.420, "dur": 11.553,
    "args": {
      "External id": 532952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622271075.837, "dur": 39.468,
    "args": {
      "External id": 532953,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271080.263, "dur": 1.476,
    "args": {
      "External id": 532954,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271084.004, "dur": 1.081,
    "args": {
      "External id": 532955,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271118.224, "dur": 14.304,
    "args": {
      "External id": 532956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271133.868, "dur": 11.831,
    "args": {
      "External id": 532957,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622271155.657, "dur": 2.631,
    "args": {
      "External id": 532958,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622271180.043, "dur": 4.554,
    "args": {
      "External id": 532959,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271182.943, "dur": 0.714,
    "args": {
      "External id": 532960,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622271251.656, "dur": 49.471,
    "args": {
      "External id": 532961,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622271305.936, "dur": 5.413,
    "args": {
      "External id": 532962,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271309.288, "dur": 0.983,
    "args": {
      "External id": 532963,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271312.806, "dur": 24.828,
    "args": {
      "External id": 532964,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622271342.616, "dur": 7.674,
    "args": {
      "External id": 532965,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622271344.166, "dur": 5.447,
    "args": {
      "External id": 532966,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271346.294, "dur": 3.112,
    "args": {
      "External id": 532967,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622271352.987, "dur": 43.404,
    "args": {
      "External id": 532968,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622271353.887, "dur": 41.760,
    "args": {
      "External id": 532969,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271400.236, "dur": 16.128,
    "args": {
      "External id": 532970,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622271421.264, "dur": 3.939,
    "args": {
      "External id": 532971,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271423.516, "dur": 0.798,
    "args": {
      "External id": 532972,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622271429.083, "dur": 49.053,
    "args": {
      "External id": 532973,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622271429.895, "dur": 5.103,
    "args": {
      "External id": 532974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622271430.764, "dur": 3.686,
    "args": {
      "External id": 532975,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271432.821, "dur": 1.266,
    "args": {
      "External id": 532976,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622271435.795, "dur": 41.885,
    "args": {
      "External id": 532977,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622271436.530, "dur": 40.666,
    "args": {
      "External id": 532978,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622271482.389, "dur": 3.533,
    "args": {
      "External id": 532979,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271484.089, "dur": 0.741,
    "args": {
      "External id": 532980,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622271491.225, "dur": 1.560,
    "args": {
      "External id": 532981,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622271500.069, "dur": 6.793,
    "args": {
      "External id": 532982,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622271502.375, "dur": 4.123,
    "args": {
      "External id": 532983,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622271586.324, "dur": 269.714,
    "args": {
      "External id": 532984,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622271588.620, "dur": 2.441,
    "args": {
      "External id": 532985,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622271595.208, "dur": 260.220,
    "args": {
      "External id": 532986,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622271596.424, "dur": 0.377,
    "args": {
      "External id": 532987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622271598.181, "dur": 23.575,
    "args": {
      "External id": 532988,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622271623.298, "dur": 6.384,
    "args": {
      "External id": 532989,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271625.856, "dur": 3.566,
    "args": {
      "External id": 532990,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622271630.719, "dur": 94.050,
    "args": {
      "External id": 532991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622271631.540, "dur": 1.434,
    "args": {
      "External id": 532992,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622271700.029, "dur": 24.359,
    "args": {
      "External id": 532993,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271705.574, "dur": 3.244,
    "args": {
      "External id": 532994,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622271726.510, "dur": 26.647,
    "args": {
      "External id": 532995,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271754.639, "dur": 20.295,
    "args": {
      "External id": 532996,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622271777.646, "dur": 15.528,
    "args": {
      "External id": 532997,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271794.337, "dur": 11.888,
    "args": {
      "External id": 532998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622271807.918, "dur": 21.066,
    "args": {
      "External id": 532999,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271809.667, "dur": 1.634,
    "args": {
      "External id": 533000,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271814.323, "dur": 0.974,
    "args": {
      "External id": 533001,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271830.907, "dur": 11.560,
    "args": {
      "External id": 533002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622271843.846, "dur": 10.418,
    "args": {
      "External id": 533003,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622271864.173, "dur": 2.708,
    "args": {
      "External id": 533004,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622271876.983, "dur": 3.833,
    "args": {
      "External id": 533005,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622271879.559, "dur": 0.368,
    "args": {
      "External id": 533006,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622271954.186, "dur": 51.860,
    "args": {
      "External id": 533007,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622272011.283, "dur": 4.831,
    "args": {
      "External id": 533008,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272014.428, "dur": 0.622,
    "args": {
      "External id": 533009,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272017.624, "dur": 24.020,
    "args": {
      "External id": 533010,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622272046.667, "dur": 7.619,
    "args": {
      "External id": 533011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622272048.868, "dur": 4.667,
    "args": {
      "External id": 533012,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272051.722, "dur": 1.587,
    "args": {
      "External id": 533013,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622272057.009, "dur": 59.669,
    "args": {
      "External id": 533014,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622272058.047, "dur": 57.714,
    "args": {
      "External id": 533015,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272122.158, "dur": 17.455,
    "args": {
      "External id": 533016,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622272145.947, "dur": 5.103,
    "args": {
      "External id": 533017,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272148.734, "dur": 1.203,
    "args": {
      "External id": 533018,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622272155.109, "dur": 52.013,
    "args": {
      "External id": 533019,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622272156.326, "dur": 4.850,
    "args": {
      "External id": 533020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622272157.625, "dur": 2.964,
    "args": {
      "External id": 533021,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272159.646, "dur": 0.803,
    "args": {
      "External id": 533022,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622272161.801, "dur": 44.995,
    "args": {
      "External id": 533023,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622272162.860, "dur": 43.472,
    "args": {
      "External id": 533024,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622272211.402, "dur": 4.286,
    "args": {
      "External id": 533025,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272213.560, "dur": 1.016,
    "args": {
      "External id": 533026,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622272221.619, "dur": 1.829,
    "args": {
      "External id": 533027,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622272231.464, "dur": 8.960,
    "args": {
      "External id": 533028,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622272234.235, "dur": 5.815,
    "args": {
      "External id": 533029,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622272331.582, "dur": 187.115,
    "args": {
      "External id": 533030,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622272333.897, "dur": 2.147,
    "args": {
      "External id": 533031,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622272337.732, "dur": 180.268,
    "args": {
      "External id": 533032,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622272339.226, "dur": 0.758,
    "args": {
      "External id": 533033,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622272341.748, "dur": 23.366,
    "args": {
      "External id": 533034,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622272366.755, "dur": 3.709,
    "args": {
      "External id": 533035,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272368.826, "dur": 1.298,
    "args": {
      "External id": 533036,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622272371.653, "dur": 23.938,
    "args": {
      "External id": 533037,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622272372.873, "dur": 1.752,
    "args": {
      "External id": 533038,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622272375.874, "dur": 19.406,
    "args": {
      "External id": 533039,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272380.981, "dur": 2.420,
    "args": {
      "External id": 533040,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622272397.222, "dur": 18.851,
    "args": {
      "External id": 533041,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272417.616, "dur": 12.813,
    "args": {
      "External id": 533042,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622272439.499, "dur": 17.947,
    "args": {
      "External id": 533043,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272459.021, "dur": 12.438,
    "args": {
      "External id": 533044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622272473.164, "dur": 19.901,
    "args": {
      "External id": 533045,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272475.105, "dur": 1.919,
    "args": {
      "External id": 533046,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272479.954, "dur": 0.873,
    "args": {
      "External id": 533047,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272494.915, "dur": 10.509,
    "args": {
      "External id": 533048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272506.541, "dur": 10.457,
    "args": {
      "External id": 533049,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622272525.245, "dur": 1.755,
    "args": {
      "External id": 533050,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622272535.364, "dur": 3.655,
    "args": {
      "External id": 533051,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272537.799, "dur": 0.413,
    "args": {
      "External id": 533052,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622272601.564, "dur": 41.158,
    "args": {
      "External id": 533053,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622272646.795, "dur": 7.050,
    "args": {
      "External id": 533054,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272649.694, "dur": 2.964,
    "args": {
      "External id": 533055,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272690.733, "dur": 26.045,
    "args": {
      "External id": 533056,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622272723.085, "dur": 6.269,
    "args": {
      "External id": 533057,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622272724.559, "dur": 3.961,
    "args": {
      "External id": 533058,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272727.332, "dur": 0.962,
    "args": {
      "External id": 533059,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622272732.223, "dur": 46.920,
    "args": {
      "External id": 533060,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622272733.374, "dur": 45.060,
    "args": {
      "External id": 533061,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622272782.934, "dur": 14.943,
    "args": {
      "External id": 533062,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622272803.982, "dur": 4.631,
    "args": {
      "External id": 533063,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272806.644, "dur": 1.049,
    "args": {
      "External id": 533064,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622272812.749, "dur": 48.855,
    "args": {
      "External id": 533065,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622272814.194, "dur": 4.609,
    "args": {
      "External id": 533066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622272815.323, "dur": 2.886,
    "args": {
      "External id": 533067,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272817.030, "dur": 1.066,
    "args": {
      "External id": 533068,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622272819.367, "dur": 41.887,
    "args": {
      "External id": 533069,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622272819.926, "dur": 40.824,
    "args": {
      "External id": 533070,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622272865.667, "dur": 4.134,
    "args": {
      "External id": 533071,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622272867.821, "dur": 0.697,
    "args": {
      "External id": 533072,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622272875.459, "dur": 1.733,
    "args": {
      "External id": 533073,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622272884.844, "dur": 7.322,
    "args": {
      "External id": 533074,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622272887.590, "dur": 4.278,
    "args": {
      "External id": 533075,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622272976.648, "dur": 217.878,
    "args": {
      "External id": 533076,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622272979.236, "dur": 5.302,
    "args": {
      "External id": 533077,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622272986.371, "dur": 207.563,
    "args": {
      "External id": 533078,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622272987.993, "dur": 0.385,
    "args": {
      "External id": 533079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622272990.105, "dur": 22.472,
    "args": {
      "External id": 533080,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622273014.209, "dur": 6.114,
    "args": {
      "External id": 533081,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273018.885, "dur": 1.191,
    "args": {
      "External id": 533082,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622273021.253, "dur": 22.195,
    "args": {
      "External id": 533083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622273022.265, "dur": 1.943,
    "args": {
      "External id": 533084,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622273025.613, "dur": 17.554,
    "args": {
      "External id": 533085,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273027.979, "dur": 2.602,
    "args": {
      "External id": 533086,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622273045.034, "dur": 23.201,
    "args": {
      "External id": 533087,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273069.551, "dur": 13.267,
    "args": {
      "External id": 533088,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622273085.504, "dur": 32.632,
    "args": {
      "External id": 533089,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273120.395, "dur": 14.431,
    "args": {
      "External id": 533090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622273136.947, "dur": 23.991,
    "args": {
      "External id": 533091,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273139.502, "dur": 2.392,
    "args": {
      "External id": 533092,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273144.023, "dur": 2.419,
    "args": {
      "External id": 533093,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273163.048, "dur": 12.849,
    "args": {
      "External id": 533094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273177.015, "dur": 15.528,
    "args": {
      "External id": 533095,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622273202.690, "dur": 2.301,
    "args": {
      "External id": 533096,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622273214.226, "dur": 4.029,
    "args": {
      "External id": 533097,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273216.738, "dur": 0.612,
    "args": {
      "External id": 533098,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622273282.179, "dur": 47.221,
    "args": {
      "External id": 533099,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622273334.001, "dur": 4.656,
    "args": {
      "External id": 533100,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273336.936, "dur": 0.679,
    "args": {
      "External id": 533101,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273339.931, "dur": 23.600,
    "args": {
      "External id": 533102,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622273367.828, "dur": 7.015,
    "args": {
      "External id": 533103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622273369.621, "dur": 4.509,
    "args": {
      "External id": 533104,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273372.021, "dur": 1.952,
    "args": {
      "External id": 533105,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622273377.274, "dur": 40.821,
    "args": {
      "External id": 533106,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622273378.843, "dur": 38.670,
    "args": {
      "External id": 533107,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273421.628, "dur": 13.571,
    "args": {
      "External id": 533108,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622273439.438, "dur": 23.120,
    "args": {
      "External id": 533109,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622273441.983, "dur": 20.202,
    "args": {
      "External id": 533110,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273447.088, "dur": 0.722,
    "args": {
      "External id": 533111,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 7270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622273467.670, "dur": 28.595,
    "args": {
      "External id": 533112,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 7271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622273469.714, "dur": 26.298,
    "args": {
      "External id": 533113,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], [], []], "Ev Idx": 7272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273474.927, "dur": 4.159,
    "args": {
      "External id": 533114,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273480.434, "dur": 15.018,
    "args": {
      "External id": 533115,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622273510.154, "dur": 6.464,
    "args": {
      "External id": 533116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 7275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622273513.080, "dur": 3.229,
    "args": {
      "External id": 533117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 7276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622273517.767, "dur": 1.300,
    "args": {
      "External id": 533118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622273518.478, "dur": 0.518,
    "args": {
      "External id": 533119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273560.801, "dur": 22.490,
    "args": {
      "External id": 533120,"Sequence number": 6319605, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 7279
    }
  },
  {
    "ph": "s", "id": 23, "pid": 4183438, "tid": 4183438, "ts": 679622273560.801,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622273589.211, "dur": 5.680,
    "args": {
      "External id": 533121,"Sequence number": 6319606, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 7280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273592.493, "dur": 1.105,
    "args": {
      "External id": 533122,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 7281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622273597.417, "dur": 6.828,
    "args": {
      "External id": 533123,"Sequence number": 6319606, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 7282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273602.027, "dur": 1.184,
    "args": {
      "External id": 533124,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "2"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 7283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622273605.467, "dur": 3.163,
    "args": {
      "External id": 533125,"Sequence number": 6319606, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 7284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273607.237, "dur": 0.798,
    "args": {
      "External id": 533126,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "2"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 7285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622273612.547, "dur": 6.359,
    "args": {
      "External id": 533127,"Sequence number": 6319606, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 7286
    }
  },
  {
    "ph": "s", "id": 22, "pid": 4183438, "tid": 4183438, "ts": 679622273612.547,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273616.345, "dur": 1.030,
    "args": {
      "External id": 533128,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 7287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622273620.030, "dur": 5.394,
    "args": {
      "External id": 533129,"Sequence number": 6319607, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 7288
    }
  },
  {
    "ph": "s", "id": 21, "pid": 4183438, "tid": 4183438, "ts": 679622273620.030,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273623.478, "dur": 1.188,
    "args": {
      "External id": 533130,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 7289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622273626.457, "dur": 6.360,
    "args": {
      "External id": 533131,"Sequence number": 6319608, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 7290
    }
  },
  {
    "ph": "s", "id": 20, "pid": 4183438, "tid": 4183438, "ts": 679622273626.457,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273630.784, "dur": 1.183,
    "args": {
      "External id": 533132,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "1024"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 7291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622273633.757, "dur": 4.946,
    "args": {
      "External id": 533133,"Sequence number": 6319609, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 7292
    }
  },
  {
    "ph": "s", "id": 19, "pid": 4183438, "tid": 4183438, "ts": 679622273633.757,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273636.865, "dur": 1.128,
    "args": {
      "External id": 533134,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "1024"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 7293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622273642.458, "dur": 74.130,
    "args": {
      "External id": 533135,"Sequence number": 6319610, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 7294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622273644.095, "dur": 72.188,
    "args": {
      "External id": 533136,"Sequence number": 6319610, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 7295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622273646.737, "dur": 7.427,
    "args": {
      "External id": 533137,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 7296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622273649.313, "dur": 4.319,
    "args": {
      "External id": 533138,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273694.937, "dur": 20.771,
    "args": {
      "External id": 533139,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 7298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622273745.029, "dur": 7.491,
    "args": {
      "External id": 533140,"Sequence number": 6319610, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 7299
    }
  },
  {
    "ph": "s", "id": 18, "pid": 4183438, "tid": 4183438, "ts": 679622273745.029,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622273755.234, "dur": 1.161,
    "args": {
      "External id": 533141,"Sequence number": 6319611, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 7300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622273789.120, "dur": 22525.703,
    "args": {
      "External id": 533142,"Sequence number": 6319611, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1], [1], [1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536], [32000, 1024], [], [], [], [], []], "Ev Idx": 7301
    }
  },
  {
    "ph": "s", "id": 17, "pid": 4183438, "tid": 4183438, "ts": 679622273789.120,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622273804.895, "dur": 29.784,
    "args": {
      "External id": 533143,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 7302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622273805.906, "dur": 28.573,
    "args": {
      "External id": 533144,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 7303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622273807.596, "dur": 7.323,
    "args": {
      "External id": 533145,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[4096, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 7304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622273810.100, "dur": 4.322,
    "args": {
      "External id": 533146,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273815.909, "dur": 17.995,
    "args": {
      "External id": 533147,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [4096, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 7306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622273850.633, "dur": 26.833,
    "args": {
      "External id": 533148,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 7307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622273851.893, "dur": 7.027,
    "args": {
      "External id": 533149,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 7308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273854.253, "dur": 4.347,
    "args": {
      "External id": 533150,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273860.360, "dur": 16.895,
    "args": {
      "External id": 533151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 7310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273862.368, "dur": 14.516,
    "args": {
      "External id": 533152,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 7311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622273881.217, "dur": 20.002,
    "args": {
      "External id": 533153,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 7312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622273881.949, "dur": 5.289,
    "args": {
      "External id": 533154,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 7313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273883.924, "dur": 3.033,
    "args": {
      "External id": 533155,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273888.040, "dur": 12.986,
    "args": {
      "External id": 533156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273888.687, "dur": 12.039,
    "args": {
      "External id": 533157,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[32000, 1024], []], "Ev Idx": 7316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622273907.023, "dur": 18.707,
    "args": {
      "External id": 533158,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 7317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622273911.400, "dur": 2.504,
    "args": {
      "External id": 533159,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273914.609, "dur": 10.863,
    "args": {
      "External id": 533160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 7319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273915.230, "dur": 9.906,
    "args": {
      "External id": 533161,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 4183438,
    "ts": 679622273931.010, "dur": 21.795,
    "args": {
      "External id": 533162,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622273955.830, "dur": 49.619,
    "args": {
      "External id": 533163,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622273958.155, "dur": 46.808,
    "args": {
      "External id": 533164,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273962.984, "dur": 1.055,
    "args": {
      "External id": 533165,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 7324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622273965.656, "dur": 23.719,
    "args": {
      "External id": 533166,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622273967.522, "dur": 21.648,
    "args": {
      "External id": 533167,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 7326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622273972.136, "dur": 2.758,
    "args": {
      "External id": 533168,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622273975.820, "dur": 12.975,
    "args": {
      "External id": 533169,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 7328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622274009.406, "dur": 16398.799,
    "args": {
      "External id": 533170,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 7329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622274010.619, "dur": 16396.493,
    "args": {
      "External id": 533171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 7330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622290418.891, "dur": 6.815,
    "args": {
      "External id": 533172,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622290423.133, "dur": 0.891,
    "args": {
      "External id": 533173,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622290430.993, "dur": 104.256,
    "args": {
      "External id": 533174,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622290432.733, "dur": 7.321,
    "args": {
      "External id": 533175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622290435.074, "dur": 4.148,
    "args": {
      "External id": 533176,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622290437.824, "dur": 1.107,
    "args": {
      "External id": 533177,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622290441.413, "dur": 92.863,
    "args": {
      "External id": 533178,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622290443.054, "dur": 90.499,
    "args": {
      "External id": 533179,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622290539.119, "dur": 4.351,
    "args": {
      "External id": 533180,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622290541.262, "dur": 0.702,
    "args": {
      "External id": 533181,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622290550.970, "dur": 2.520,
    "args": {
      "External id": 533182,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622290562.372, "dur": 6.756,
    "args": {
      "External id": 533183,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622290564.709, "dur": 4.110,
    "args": {
      "External id": 533184,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622290731.287, "dur": 228.132,
    "args": {
      "External id": 533185,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622290735.176, "dur": 3.069,
    "args": {
      "External id": 533186,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622290742.758, "dur": 216.093,
    "args": {
      "External id": 533187,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622290744.694, "dur": 0.458,
    "args": {
      "External id": 533188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622290747.130, "dur": 27.562,
    "args": {
      "External id": 533189,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622290776.768, "dur": 5.265,
    "args": {
      "External id": 533190,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622290778.854, "dur": 2.811,
    "args": {
      "External id": 533191,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622290783.303, "dur": 40.916,
    "args": {
      "External id": 533192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622290798.814, "dur": 1.823,
    "args": {
      "External id": 533193,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622290802.150, "dur": 21.700,
    "args": {
      "External id": 533194,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622290806.402, "dur": 3.171,
    "args": {
      "External id": 533195,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622290826.204, "dur": 22.999,
    "args": {
      "External id": 533196,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622290850.957, "dur": 18.483,
    "args": {
      "External id": 533197,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622290872.484, "dur": 15.521,
    "args": {
      "External id": 533198,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622290889.463, "dur": 15.566,
    "args": {
      "External id": 533199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622290906.895, "dur": 21.994,
    "args": {
      "External id": 533200,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622290908.782, "dur": 1.938,
    "args": {
      "External id": 533201,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622290913.145, "dur": 0.705,
    "args": {
      "External id": 533202,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622290930.636, "dur": 13.890,
    "args": {
      "External id": 533203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622290946.329, "dur": 11.624,
    "args": {
      "External id": 533204,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622290967.267, "dur": 2.371,
    "args": {
      "External id": 533205,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622290976.900, "dur": 4.547,
    "args": {
      "External id": 533206,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622290979.588, "dur": 0.752,
    "args": {
      "External id": 533207,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622291061.482, "dur": 75.801,
    "args": {
      "External id": 533208,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622291144.545, "dur": 7.617,
    "args": {
      "External id": 533209,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291147.865, "dur": 1.097,
    "args": {
      "External id": 533210,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291153.862, "dur": 28.409,
    "args": {
      "External id": 533211,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622291188.615, "dur": 7.382,
    "args": {
      "External id": 533212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622291191.097, "dur": 4.177,
    "args": {
      "External id": 533213,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291193.527, "dur": 1.504,
    "args": {
      "External id": 533214,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622291199.975, "dur": 52.107,
    "args": {
      "External id": 533215,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622291201.669, "dur": 49.777,
    "args": {
      "External id": 533216,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291257.023, "dur": 17.725,
    "args": {
      "External id": 533217,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622291281.259, "dur": 4.790,
    "args": {
      "External id": 533218,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291284.192, "dur": 1.009,
    "args": {
      "External id": 533219,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622291290.509, "dur": 53.483,
    "args": {
      "External id": 533220,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622291291.733, "dur": 4.575,
    "args": {
      "External id": 533221,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622291292.662, "dur": 3.054,
    "args": {
      "External id": 533222,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291294.591, "dur": 0.968,
    "args": {
      "External id": 533223,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622291297.297, "dur": 46.294,
    "args": {
      "External id": 533224,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622291298.388, "dur": 44.722,
    "args": {
      "External id": 533225,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622291348.199, "dur": 4.245,
    "args": {
      "External id": 533226,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291350.328, "dur": 0.856,
    "args": {
      "External id": 533227,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622291359.352, "dur": 2.090,
    "args": {
      "External id": 533228,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622291369.218, "dur": 9.648,
    "args": {
      "External id": 533229,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622291371.664, "dur": 6.864,
    "args": {
      "External id": 533230,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622291477.211, "dur": 328.678,
    "args": {
      "External id": 533231,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622291479.896, "dur": 2.678,
    "args": {
      "External id": 533232,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622291484.602, "dur": 320.789,
    "args": {
      "External id": 533233,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622291486.472, "dur": 0.304,
    "args": {
      "External id": 533234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622291488.070, "dur": 25.227,
    "args": {
      "External id": 533235,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622291515.080, "dur": 3.262,
    "args": {
      "External id": 533236,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291517.060, "dur": 1.033,
    "args": {
      "External id": 533237,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622291519.458, "dur": 28.712,
    "args": {
      "External id": 533238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622291520.802, "dur": 2.085,
    "args": {
      "External id": 533239,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622291524.285, "dur": 23.535,
    "args": {
      "External id": 533240,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291529.550, "dur": 2.803,
    "args": {
      "External id": 533241,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622291550.036, "dur": 33.142,
    "args": {
      "External id": 533242,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291584.729, "dur": 41.571,
    "args": {
      "External id": 533243,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622291629.032, "dur": 59.374,
    "args": {
      "External id": 533244,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291691.000, "dur": 35.211,
    "args": {
      "External id": 533245,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622291728.368, "dur": 32.203,
    "args": {
      "External id": 533246,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291730.816, "dur": 2.398,
    "args": {
      "External id": 533247,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291735.808, "dur": 0.933,
    "args": {
      "External id": 533248,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291762.490, "dur": 25.940,
    "args": {
      "External id": 533249,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291789.951, "dur": 14.369,
    "args": {
      "External id": 533250,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622291814.153, "dur": 2.558,
    "args": {
      "External id": 533251,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622291826.839, "dur": 4.019,
    "args": {
      "External id": 533252,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291829.427, "dur": 0.458,
    "args": {
      "External id": 533253,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622291903.905, "dur": 48.956,
    "args": {
      "External id": 533254,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622291957.623, "dur": 5.207,
    "args": {
      "External id": 533255,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291960.453, "dur": 1.260,
    "args": {
      "External id": 533256,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622291964.573, "dur": 24.276,
    "args": {
      "External id": 533257,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622291993.649, "dur": 6.097,
    "args": {
      "External id": 533258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622291995.328, "dur": 3.774,
    "args": {
      "External id": 533259,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622291997.864, "dur": 1.030,
    "args": {
      "External id": 533260,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622292002.280, "dur": 43.300,
    "args": {
      "External id": 533261,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622292003.498, "dur": 41.404,
    "args": {
      "External id": 533262,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292049.301, "dur": 14.792,
    "args": {
      "External id": 533263,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622292069.743, "dur": 3.959,
    "args": {
      "External id": 533264,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292072.362, "dur": 0.576,
    "args": {
      "External id": 533265,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622292077.704, "dur": 90.000,
    "args": {
      "External id": 533266,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622292078.881, "dur": 6.972,
    "args": {
      "External id": 533267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622292079.862, "dur": 5.413,
    "args": {
      "External id": 533268,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292081.580, "dur": 3.475,
    "args": {
      "External id": 533269,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622292113.343, "dur": 53.449,
    "args": {
      "External id": 533270,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622292119.030, "dur": 46.842,
    "args": {
      "External id": 533271,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622292173.420, "dur": 4.642,
    "args": {
      "External id": 533272,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292175.657, "dur": 0.980,
    "args": {
      "External id": 533273,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622292184.738, "dur": 1.819,
    "args": {
      "External id": 533274,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292195.416, "dur": 7.152,
    "args": {
      "External id": 533275,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292198.041, "dur": 4.210,
    "args": {
      "External id": 533276,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622292291.152, "dur": 184.179,
    "args": {
      "External id": 533277,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292295.985, "dur": 2.609,
    "args": {
      "External id": 533278,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622292300.458, "dur": 174.231,
    "args": {
      "External id": 533279,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622292301.990, "dur": 0.338,
    "args": {
      "External id": 533280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622292303.787, "dur": 21.555,
    "args": {
      "External id": 533281,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622292326.983, "dur": 4.821,
    "args": {
      "External id": 533282,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292330.816, "dur": 0.718,
    "args": {
      "External id": 533283,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622292332.589, "dur": 20.995,
    "args": {
      "External id": 533284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292333.742, "dur": 1.448,
    "args": {
      "External id": 533285,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622292336.483, "dur": 16.773,
    "args": {
      "External id": 533286,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292339.315, "dur": 2.538,
    "args": {
      "External id": 533287,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622292355.582, "dur": 21.000,
    "args": {
      "External id": 533288,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292378.002, "dur": 13.566,
    "args": {
      "External id": 533289,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622292393.979, "dur": 13.814,
    "args": {
      "External id": 533290,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292409.242, "dur": 12.411,
    "args": {
      "External id": 533291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622292423.621, "dur": 21.003,
    "args": {
      "External id": 533292,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292427.758, "dur": 1.737,
    "args": {
      "External id": 533293,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292431.661, "dur": 0.704,
    "args": {
      "External id": 533294,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292446.585, "dur": 11.409,
    "args": {
      "External id": 533295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292459.220, "dur": 14.343,
    "args": {
      "External id": 533296,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622292484.992, "dur": 1.933,
    "args": {
      "External id": 533297,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622292495.542, "dur": 3.834,
    "args": {
      "External id": 533298,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292497.758, "dur": 0.566,
    "args": {
      "External id": 533299,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622292565.157, "dur": 40.374,
    "args": {
      "External id": 533300,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622292609.889, "dur": 5.064,
    "args": {
      "External id": 533301,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292612.882, "dur": 1.059,
    "args": {
      "External id": 533302,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292616.306, "dur": 21.493,
    "args": {
      "External id": 533303,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622292642.775, "dur": 44.430,
    "args": {
      "External id": 533304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622292644.198, "dur": 6.037,
    "args": {
      "External id": 533305,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292646.602, "dur": 3.411,
    "args": {
      "External id": 533306,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622292691.749, "dur": 50.052,
    "args": {
      "External id": 533307,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622292692.983, "dur": 48.200,
    "args": {
      "External id": 533308,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292746.062, "dur": 15.701,
    "args": {
      "External id": 533309,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622292768.099, "dur": 5.145,
    "args": {
      "External id": 533310,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292771.196, "dur": 1.076,
    "args": {
      "External id": 533311,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622292777.082, "dur": 49.370,
    "args": {
      "External id": 533312,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622292778.163, "dur": 5.082,
    "args": {
      "External id": 533313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622292779.577, "dur": 3.102,
    "args": {
      "External id": 533314,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292781.632, "dur": 0.877,
    "args": {
      "External id": 533315,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622292783.841, "dur": 42.157,
    "args": {
      "External id": 533316,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622292784.283, "dur": 41.243,
    "args": {
      "External id": 533317,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622292830.699, "dur": 4.409,
    "args": {
      "External id": 533318,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292832.955, "dur": 0.984,
    "args": {
      "External id": 533319,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622292841.041, "dur": 1.526,
    "args": {
      "External id": 533320,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292850.429, "dur": 7.605,
    "args": {
      "External id": 533321,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292852.753, "dur": 4.974,
    "args": {
      "External id": 533322,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622292943.289, "dur": 202.134,
    "args": {
      "External id": 533323,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292945.912, "dur": 2.342,
    "args": {
      "External id": 533324,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622292952.155, "dur": 192.741,
    "args": {
      "External id": 533325,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622292953.422, "dur": 0.296,
    "args": {
      "External id": 533326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622292955.264, "dur": 20.543,
    "args": {
      "External id": 533327,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622292977.154, "dur": 5.999,
    "args": {
      "External id": 533328,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622292979.649, "dur": 3.174,
    "args": {
      "External id": 533329,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622292984.205, "dur": 21.668,
    "args": {
      "External id": 533330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622292985.384, "dur": 1.895,
    "args": {
      "External id": 533331,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622292988.645, "dur": 16.980,
    "args": {
      "External id": 533332,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622292991.531, "dur": 2.365,
    "args": {
      "External id": 533333,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622293007.463, "dur": 19.159,
    "args": {
      "External id": 533334,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293028.138, "dur": 13.589,
    "args": {
      "External id": 533335,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622293044.551, "dur": 14.857,
    "args": {
      "External id": 533336,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293060.809, "dur": 13.218,
    "args": {
      "External id": 533337,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622293075.851, "dur": 39.551,
    "args": {
      "External id": 533338,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293077.628, "dur": 1.692,
    "args": {
      "External id": 533339,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293081.865, "dur": 2.780,
    "args": {
      "External id": 533340,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293118.261, "dur": 14.059,
    "args": {
      "External id": 533341,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293133.619, "dur": 10.158,
    "args": {
      "External id": 533342,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622293152.966, "dur": 2.338,
    "args": {
      "External id": 533343,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622293164.542, "dur": 3.804,
    "args": {
      "External id": 533344,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293166.953, "dur": 0.447,
    "args": {
      "External id": 533345,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622293233.717, "dur": 44.573,
    "args": {
      "External id": 533346,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622293282.807, "dur": 5.030,
    "args": {
      "External id": 533347,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293285.674, "dur": 1.229,
    "args": {
      "External id": 533348,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293289.161, "dur": 22.081,
    "args": {
      "External id": 533349,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622293315.954, "dur": 6.606,
    "args": {
      "External id": 533350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622293317.668, "dur": 4.241,
    "args": {
      "External id": 533351,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293320.153, "dur": 1.470,
    "args": {
      "External id": 533352,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622293324.959, "dur": 41.756,
    "args": {
      "External id": 533353,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622293326.150, "dur": 39.947,
    "args": {
      "External id": 533354,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293370.628, "dur": 14.380,
    "args": {
      "External id": 533355,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622293390.417, "dur": 3.914,
    "args": {
      "External id": 533356,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293392.774, "dur": 0.634,
    "args": {
      "External id": 533357,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622293397.697, "dur": 47.060,
    "args": {
      "External id": 533358,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622293398.974, "dur": 4.195,
    "args": {
      "External id": 533359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622293400.025, "dur": 2.620,
    "args": {
      "External id": 533360,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293401.960, "dur": 0.537,
    "args": {
      "External id": 533361,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622293403.894, "dur": 40.568,
    "args": {
      "External id": 533362,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622293404.727, "dur": 39.224,
    "args": {
      "External id": 533363,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622293448.678, "dur": 3.829,
    "args": {
      "External id": 533364,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293450.526, "dur": 0.749,
    "args": {
      "External id": 533365,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622293457.390, "dur": 1.279,
    "args": {
      "External id": 533366,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622293466.534, "dur": 6.870,
    "args": {
      "External id": 533367,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622293468.677, "dur": 4.437,
    "args": {
      "External id": 533368,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622293550.337, "dur": 245.530,
    "args": {
      "External id": 533369,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622293555.454, "dur": 2.320,
    "args": {
      "External id": 533370,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622293559.408, "dur": 235.853,
    "args": {
      "External id": 533371,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622293560.637, "dur": 0.560,
    "args": {
      "External id": 533372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622293562.495, "dur": 22.548,
    "args": {
      "External id": 533373,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622293586.701, "dur": 5.081,
    "args": {
      "External id": 533374,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293590.749, "dur": 0.783,
    "args": {
      "External id": 533375,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622293592.717, "dur": 22.690,
    "args": {
      "External id": 533376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622293594.380, "dur": 1.405,
    "args": {
      "External id": 533377,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622293597.285, "dur": 17.848,
    "args": {
      "External id": 533378,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293599.865, "dur": 2.509,
    "args": {
      "External id": 533379,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622293617.035, "dur": 19.831,
    "args": {
      "External id": 533380,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293638.389, "dur": 67.214,
    "args": {
      "External id": 533381,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622293710.016, "dur": 17.236,
    "args": {
      "External id": 533382,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293728.857, "dur": 13.410,
    "args": {
      "External id": 533383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622293744.483, "dur": 23.450,
    "args": {
      "External id": 533384,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293746.612, "dur": 2.281,
    "args": {
      "External id": 533385,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293751.152, "dur": 0.782,
    "args": {
      "External id": 533386,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293769.262, "dur": 12.309,
    "args": {
      "External id": 533387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293782.822, "dur": 11.427,
    "args": {
      "External id": 533388,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622293803.973, "dur": 2.344,
    "args": {
      "External id": 533389,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622293815.800, "dur": 3.793,
    "args": {
      "External id": 533390,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293818.319, "dur": 0.355,
    "args": {
      "External id": 533391,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622293892.584, "dur": 49.001,
    "args": {
      "External id": 533392,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622293946.507, "dur": 4.740,
    "args": {
      "External id": 533393,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293949.319, "dur": 0.855,
    "args": {
      "External id": 533394,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622293952.469, "dur": 24.060,
    "args": {
      "External id": 533395,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622293981.406, "dur": 6.421,
    "args": {
      "External id": 533396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622293983.421, "dur": 3.797,
    "args": {
      "External id": 533397,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622293985.680, "dur": 1.326,
    "args": {
      "External id": 533398,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622293990.626, "dur": 45.225,
    "args": {
      "External id": 533399,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622293991.782, "dur": 43.166,
    "args": {
      "External id": 533400,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294039.612, "dur": 22.119,
    "args": {
      "External id": 533401,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622294067.610, "dur": 3.976,
    "args": {
      "External id": 533402,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294069.929, "dur": 0.853,
    "args": {
      "External id": 533403,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622294075.627, "dur": 69.515,
    "args": {
      "External id": 533404,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622294076.992, "dur": 4.656,
    "args": {
      "External id": 533405,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622294077.968, "dur": 3.124,
    "args": {
      "External id": 533406,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294080.051, "dur": 0.808,
    "args": {
      "External id": 533407,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622294082.357, "dur": 62.339,
    "args": {
      "External id": 533408,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622294083.031, "dur": 60.641,
    "args": {
      "External id": 533409,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622294150.836, "dur": 4.626,
    "args": {
      "External id": 533410,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294153.326, "dur": 0.813,
    "args": {
      "External id": 533411,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622294161.580, "dur": 1.637,
    "args": {
      "External id": 533412,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622294171.808, "dur": 9.909,
    "args": {
      "External id": 533413,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622294174.198, "dur": 7.117,
    "args": {
      "External id": 533414,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622294270.060, "dur": 313.098,
    "args": {
      "External id": 533415,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622294272.454, "dur": 2.189,
    "args": {
      "External id": 533416,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622294276.214, "dur": 306.552,
    "args": {
      "External id": 533417,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622294278.080, "dur": 0.421,
    "args": {
      "External id": 533418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622294280.080, "dur": 28.528,
    "args": {
      "External id": 533419,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622294310.272, "dur": 3.535,
    "args": {
      "External id": 533420,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294312.691, "dur": 0.846,
    "args": {
      "External id": 533421,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622294314.867, "dur": 34.847,
    "args": {
      "External id": 533422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622294315.868, "dur": 1.737,
    "args": {
      "External id": 533423,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622294319.316, "dur": 30.048,
    "args": {
      "External id": 533424,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294325.201, "dur": 2.476,
    "args": {
      "External id": 533425,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622294351.448, "dur": 40.637,
    "args": {
      "External id": 533426,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294393.620, "dur": 41.800,
    "args": {
      "External id": 533427,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622294438.149, "dur": 43.050,
    "args": {
      "External id": 533428,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294482.676, "dur": 31.520,
    "args": {
      "External id": 533429,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622294515.900, "dur": 28.136,
    "args": {
      "External id": 533430,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294517.747, "dur": 1.950,
    "args": {
      "External id": 533431,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294522.043, "dur": 0.864,
    "args": {
      "External id": 533432,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294545.600, "dur": 22.822,
    "args": {
      "External id": 533433,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294569.624, "dur": 12.182,
    "args": {
      "External id": 533434,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622294589.567, "dur": 1.913,
    "args": {
      "External id": 533435,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622294599.305, "dur": 3.555,
    "args": {
      "External id": 533436,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294601.591, "dur": 0.328,
    "args": {
      "External id": 533437,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622294719.477, "dur": 52.536,
    "args": {
      "External id": 533438,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622294777.740, "dur": 8.241,
    "args": {
      "External id": 533439,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294781.287, "dur": 3.031,
    "args": {
      "External id": 533440,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294787.420, "dur": 24.299,
    "args": {
      "External id": 533441,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622294817.178, "dur": 6.486,
    "args": {
      "External id": 533442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622294818.913, "dur": 3.846,
    "args": {
      "External id": 533443,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294821.319, "dur": 0.981,
    "args": {
      "External id": 533444,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622294826.506, "dur": 43.473,
    "args": {
      "External id": 533445,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622294827.724, "dur": 41.597,
    "args": {
      "External id": 533446,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622294873.825, "dur": 14.679,
    "args": {
      "External id": 533447,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622294894.173, "dur": 3.802,
    "args": {
      "External id": 533448,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294896.397, "dur": 0.694,
    "args": {
      "External id": 533449,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622294901.757, "dur": 47.267,
    "args": {
      "External id": 533450,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622294902.851, "dur": 4.630,
    "args": {
      "External id": 533451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622294903.974, "dur": 2.998,
    "args": {
      "External id": 533452,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294905.859, "dur": 0.835,
    "args": {
      "External id": 533453,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622294908.180, "dur": 40.512,
    "args": {
      "External id": 533454,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622294908.944, "dur": 39.206,
    "args": {
      "External id": 533455,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622294953.219, "dur": 4.069,
    "args": {
      "External id": 533456,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622294955.109, "dur": 1.022,
    "args": {
      "External id": 533457,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622294963.568, "dur": 1.550,
    "args": {
      "External id": 533458,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622294973.321, "dur": 12.734,
    "args": {
      "External id": 533459,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622294979.627, "dur": 6.063,
    "args": {
      "External id": 533460,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622295072.486, "dur": 199.080,
    "args": {
      "External id": 533461,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622295074.700, "dur": 2.416,
    "args": {
      "External id": 533462,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622295078.892, "dur": 192.168,
    "args": {
      "External id": 533463,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622295080.243, "dur": 0.550,
    "args": {
      "External id": 533464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622295082.305, "dur": 40.202,
    "args": {
      "External id": 533465,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622295125.337, "dur": 3.653,
    "args": {
      "External id": 533466,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295127.513, "dur": 1.197,
    "args": {
      "External id": 533467,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622295130.012, "dur": 26.286,
    "args": {
      "External id": 533468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622295131.687, "dur": 2.247,
    "args": {
      "External id": 533469,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622295135.188, "dur": 20.788,
    "args": {
      "External id": 533470,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295140.425, "dur": 2.586,
    "args": {
      "External id": 533471,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622295157.884, "dur": 19.444,
    "args": {
      "External id": 533472,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295178.690, "dur": 12.684,
    "args": {
      "External id": 533473,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622295193.864, "dur": 13.689,
    "args": {
      "External id": 533474,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295208.884, "dur": 11.377,
    "args": {
      "External id": 533475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622295222.163, "dur": 19.429,
    "args": {
      "External id": 533476,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295224.303, "dur": 1.500,
    "args": {
      "External id": 533477,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295228.163, "dur": 0.913,
    "args": {
      "External id": 533478,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295242.900, "dur": 13.034,
    "args": {
      "External id": 533479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295257.143, "dur": 12.898,
    "args": {
      "External id": 533480,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622295279.034, "dur": 2.576,
    "args": {
      "External id": 533481,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622295290.486, "dur": 4.008,
    "args": {
      "External id": 533482,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295293.103, "dur": 0.498,
    "args": {
      "External id": 533483,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622295359.530, "dur": 43.199,
    "args": {
      "External id": 533484,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622295407.021, "dur": 4.934,
    "args": {
      "External id": 533485,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295409.656, "dur": 1.261,
    "args": {
      "External id": 533486,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295413.161, "dur": 21.197,
    "args": {
      "External id": 533487,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622295439.006, "dur": 5.962,
    "args": {
      "External id": 533488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622295440.390, "dur": 3.767,
    "args": {
      "External id": 533489,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295442.998, "dur": 0.964,
    "args": {
      "External id": 533490,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622295447.253, "dur": 40.238,
    "args": {
      "External id": 533491,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622295448.540, "dur": 38.303,
    "args": {
      "External id": 533492,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295490.876, "dur": 14.095,
    "args": {
      "External id": 533493,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622295510.646, "dur": 4.378,
    "args": {
      "External id": 533494,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295513.263, "dur": 0.920,
    "args": {
      "External id": 533495,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622295518.779, "dur": 49.916,
    "args": {
      "External id": 533496,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622295519.959, "dur": 6.688,
    "args": {
      "External id": 533497,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622295521.008, "dur": 5.099,
    "args": {
      "External id": 533498,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295523.085, "dur": 2.852,
    "args": {
      "External id": 533499,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622295527.502, "dur": 40.833,
    "args": {
      "External id": 533500,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622295528.346, "dur": 39.399,
    "args": {
      "External id": 533501,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622295572.414, "dur": 3.827,
    "args": {
      "External id": 533502,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295574.168, "dur": 1.034,
    "args": {
      "External id": 533503,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622295581.192, "dur": 1.290,
    "args": {
      "External id": 533504,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622295611.274, "dur": 8.530,
    "args": {
      "External id": 533505,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622295615.463, "dur": 4.023,
    "args": {
      "External id": 533506,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622295743.204, "dur": 184.055,
    "args": {
      "External id": 533507,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622295747.963, "dur": 3.570,
    "args": {
      "External id": 533508,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622295753.183, "dur": 173.590,
    "args": {
      "External id": 533509,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622295754.799, "dur": 0.329,
    "args": {
      "External id": 533510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622295756.613, "dur": 22.073,
    "args": {
      "External id": 533511,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622295780.381, "dur": 5.833,
    "args": {
      "External id": 533512,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295784.976, "dur": 0.931,
    "args": {
      "External id": 533513,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622295787.176, "dur": 21.437,
    "args": {
      "External id": 533514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622295788.136, "dur": 1.380,
    "args": {
      "External id": 533515,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622295790.515, "dur": 17.738,
    "args": {
      "External id": 533516,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295793.420, "dur": 2.884,
    "args": {
      "External id": 533517,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622295810.434, "dur": 20.410,
    "args": {
      "External id": 533518,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295832.296, "dur": 14.929,
    "args": {
      "External id": 533519,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622295850.082, "dur": 14.027,
    "args": {
      "External id": 533520,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295865.434, "dur": 12.214,
    "args": {
      "External id": 533521,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622295879.518, "dur": 18.722,
    "args": {
      "External id": 533522,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295881.114, "dur": 1.765,
    "args": {
      "External id": 533523,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295885.102, "dur": 0.692,
    "args": {
      "External id": 533524,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295899.660, "dur": 12.685,
    "args": {
      "External id": 533525,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622295913.588, "dur": 12.177,
    "args": {
      "External id": 533526,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622295934.239, "dur": 2.049,
    "args": {
      "External id": 533527,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622295945.581, "dur": 4.132,
    "args": {
      "External id": 533528,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622295948.436, "dur": 0.298,
    "args": {
      "External id": 533529,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622296013.492, "dur": 48.656,
    "args": {
      "External id": 533530,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622296066.691, "dur": 5.253,
    "args": {
      "External id": 533531,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296069.933, "dur": 0.942,
    "args": {
      "External id": 533532,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296073.151, "dur": 23.527,
    "args": {
      "External id": 533533,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622296118.495, "dur": 7.322,
    "args": {
      "External id": 533534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622296120.317, "dur": 4.449,
    "args": {
      "External id": 533535,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296122.907, "dur": 1.574,
    "args": {
      "External id": 533536,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622296128.977, "dur": 50.210,
    "args": {
      "External id": 533537,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622296129.983, "dur": 48.594,
    "args": {
      "External id": 533538,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296182.842, "dur": 16.884,
    "args": {
      "External id": 533539,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622296204.559, "dur": 27.841,
    "args": {
      "External id": 533540,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622296207.197, "dur": 24.807,
    "args": {
      "External id": 533541,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296212.305, "dur": 3.334,
    "args": {
      "External id": 533542,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 7701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622296237.233, "dur": 32.227,
    "args": {
      "External id": 533543,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 7702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622296243.785, "dur": 25.423,
    "args": {
      "External id": 533544,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], [], []], "Ev Idx": 7703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296249.063, "dur": 3.935,
    "args": {
      "External id": 533545,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296254.226, "dur": 14.405,
    "args": {
      "External id": 533546,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622296282.816, "dur": 5.256,
    "args": {
      "External id": 533547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 7706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622296285.143, "dur": 2.654,
    "args": {
      "External id": 533548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 7707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622296289.130, "dur": 1.258,
    "args": {
      "External id": 533549,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622296289.843, "dur": 0.474,
    "args": {
      "External id": 533550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296332.028, "dur": 23.559,
    "args": {
      "External id": 533551,"Sequence number": 6319612, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 7710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296357.442, "dur": 14.300,
    "args": {
      "External id": 533552,"Sequence number": 6319613, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 7711
    }
  },
  {
    "ph": "s", "id": 16, "pid": 4183438, "tid": 4183438, "ts": 679622296357.442,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622296377.939, "dur": 5.695,
    "args": {
      "External id": 533553,"Sequence number": 6319614, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 7712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296381.272, "dur": 1.069,
    "args": {
      "External id": 533554,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 7713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622296386.005, "dur": 6.763,
    "args": {
      "External id": 533555,"Sequence number": 6319614, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 7714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296390.441, "dur": 0.884,
    "args": {
      "External id": 533556,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "3"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 7715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622296394.042, "dur": 2.870,
    "args": {
      "External id": 533557,"Sequence number": 6319614, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 7716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296395.881, "dur": 0.486,
    "args": {
      "External id": 533558,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "3"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 7717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622296400.970, "dur": 6.368,
    "args": {
      "External id": 533559,"Sequence number": 6319614, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 7718
    }
  },
  {
    "ph": "s", "id": 15, "pid": 4183438, "tid": 4183438, "ts": 679622296400.970,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296404.823, "dur": 1.125,
    "args": {
      "External id": 533560,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 7719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622296408.343, "dur": 5.137,
    "args": {
      "External id": 533561,"Sequence number": 6319615, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 7720
    }
  },
  {
    "ph": "s", "id": 14, "pid": 4183438, "tid": 4183438, "ts": 679622296408.343,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296412.095, "dur": 0.657,
    "args": {
      "External id": 533562,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 7721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622296414.353, "dur": 5.168,
    "args": {
      "External id": 533563,"Sequence number": 6319616, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 7722
    }
  },
  {
    "ph": "s", "id": 13, "pid": 4183438, "tid": 4183438, "ts": 679622296414.353,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296417.816, "dur": 0.658,
    "args": {
      "External id": 533564,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 7723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622296420.596, "dur": 4.924,
    "args": {
      "External id": 533565,"Sequence number": 6319617, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 7724
    }
  },
  {
    "ph": "s", "id": 12, "pid": 4183438, "tid": 4183438, "ts": 679622296420.596,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296423.879, "dur": 0.855,
    "args": {
      "External id": 533566,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 7725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622296428.967, "dur": 30.436,
    "args": {
      "External id": 533567,"Sequence number": 6319618, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 7726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622296430.527, "dur": 28.629,
    "args": {
      "External id": 533568,"Sequence number": 6319618, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 7727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622296433.049, "dur": 6.623,
    "args": {
      "External id": 533569,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 7728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622296435.274, "dur": 3.853,
    "args": {
      "External id": 533570,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296440.348, "dur": 18.282,
    "args": {
      "External id": 533571,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 7730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622296485.139, "dur": 3.919,
    "args": {
      "External id": 533572,"Sequence number": 6319618, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 7731
    }
  },
  {
    "ph": "s", "id": 11, "pid": 4183438, "tid": 4183438, "ts": 679622296485.139,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622296491.463, "dur": 0.985,
    "args": {
      "External id": 533573,"Sequence number": 6319619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 7732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622296527.048, "dur": 23448.161,
    "args": {
      "External id": 533574,"Sequence number": 6319619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1], [1], [1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536], [32000, 1024], [], [], [], [], []], "Ev Idx": 7733
    }
  },
  {
    "ph": "s", "id": 10, "pid": 4183438, "tid": 4183438, "ts": 679622296527.048,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622296543.014, "dur": 24.780,
    "args": {
      "External id": 533575,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 7734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622296543.946, "dur": 23.635,
    "args": {
      "External id": 533576,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 7735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622296545.009, "dur": 6.062,
    "args": {
      "External id": 533577,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[4096, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 7736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622296546.672, "dur": 3.888,
    "args": {
      "External id": 533578,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296551.801, "dur": 15.269,
    "args": {
      "External id": 533579,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [4096, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 7738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622296583.447, "dur": 26.665,
    "args": {
      "External id": 533580,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 7739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622296584.487, "dur": 6.963,
    "args": {
      "External id": 533581,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 7740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296587.356, "dur": 3.783,
    "args": {
      "External id": 533582,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296592.756, "dur": 17.146,
    "args": {
      "External id": 533583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 7742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296594.346, "dur": 15.061,
    "args": {
      "External id": 533584,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 7743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622296613.198, "dur": 21.130,
    "args": {
      "External id": 533585,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 7744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622296614.674, "dur": 4.617,
    "args": {
      "External id": 533586,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 7745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296616.539, "dur": 2.524,
    "args": {
      "External id": 533587,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296619.942, "dur": 14.200,
    "args": {
      "External id": 533588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296620.867, "dur": 12.942,
    "args": {
      "External id": 533589,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[32000, 1024], []], "Ev Idx": 7748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622296639.988, "dur": 60.065,
    "args": {
      "External id": 533590,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 7749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622296641.903, "dur": 2.289,
    "args": {
      "External id": 533591,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296644.789, "dur": 54.984,
    "args": {
      "External id": 533592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 7751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296647.592, "dur": 51.421,
    "args": {
      "External id": 533593,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 4183438,
    "ts": 679622296706.810, "dur": 24.669,
    "args": {
      "External id": 533594,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622296734.262, "dur": 49.126,
    "args": {
      "External id": 533595,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 7754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622296736.759, "dur": 46.159,
    "args": {
      "External id": 533596,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296741.238, "dur": 1.492,
    "args": {
      "External id": 533597,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 7756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622296743.974, "dur": 23.718,
    "args": {
      "External id": 533598,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622296745.906, "dur": 21.533,
    "args": {
      "External id": 533599,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 7758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622296751.046, "dur": 2.982,
    "args": {
      "External id": 533600,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622296754.839, "dur": 12.256,
    "args": {
      "External id": 533601,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 7760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622296787.478, "dur": 17598.379,
    "args": {
      "External id": 533602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 7761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622296789.080, "dur": 17595.606,
    "args": {
      "External id": 533603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 7762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622314395.815, "dur": 6.471,
    "args": {
      "External id": 533604,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622314399.525, "dur": 0.929,
    "args": {
      "External id": 533605,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622314407.586, "dur": 95.462,
    "args": {
      "External id": 533606,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622314409.257, "dur": 6.450,
    "args": {
      "External id": 533607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622314411.567, "dur": 3.250,
    "args": {
      "External id": 533608,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622314413.507, "dur": 1.069,
    "args": {
      "External id": 533609,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622314416.917, "dur": 85.446,
    "args": {
      "External id": 533610,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622314419.127, "dur": 82.569,
    "args": {
      "External id": 533611,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622314506.559, "dur": 4.255,
    "args": {
      "External id": 533612,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622314508.742, "dur": 0.690,
    "args": {
      "External id": 533613,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622314517.998, "dur": 2.631,
    "args": {
      "External id": 533614,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622314529.495, "dur": 7.347,
    "args": {
      "External id": 533615,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622314532.087, "dur": 4.467,
    "args": {
      "External id": 533616,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622314699.458, "dur": 191.391,
    "args": {
      "External id": 533617,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622314702.802, "dur": 3.603,
    "args": {
      "External id": 533618,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622314708.887, "dur": 181.444,
    "args": {
      "External id": 533619,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622314710.712, "dur": 0.488,
    "args": {
      "External id": 533620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622314715.284, "dur": 24.428,
    "args": {
      "External id": 533621,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622314741.899, "dur": 3.588,
    "args": {
      "External id": 533622,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622314744.334, "dur": 0.826,
    "args": {
      "External id": 533623,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622314746.639, "dur": 25.273,
    "args": {
      "External id": 533624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622314747.826, "dur": 3.600,
    "args": {
      "External id": 533625,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622314753.079, "dur": 18.387,
    "args": {
      "External id": 533626,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622314756.689, "dur": 2.894,
    "args": {
      "External id": 533627,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622314773.638, "dur": 22.060,
    "args": {
      "External id": 533628,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622314797.280, "dur": 13.107,
    "args": {
      "External id": 533629,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622314813.044, "dur": 15.073,
    "args": {
      "External id": 533630,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622314829.470, "dur": 12.972,
    "args": {
      "External id": 533631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622314844.427, "dur": 20.046,
    "args": {
      "External id": 533632,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622314846.403, "dur": 2.107,
    "args": {
      "External id": 533633,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622314851.023, "dur": 0.826,
    "args": {
      "External id": 533634,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622314866.148, "dur": 11.392,
    "args": {
      "External id": 533635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622314878.804, "dur": 10.652,
    "args": {
      "External id": 533636,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622314898.264, "dur": 2.425,
    "args": {
      "External id": 533637,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622314907.697, "dur": 4.225,
    "args": {
      "External id": 533638,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622314910.182, "dur": 0.746,
    "args": {
      "External id": 533639,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622314988.608, "dur": 55.680,
    "args": {
      "External id": 533640,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622315049.413, "dur": 6.262,
    "args": {
      "External id": 533641,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315052.108, "dur": 0.946,
    "args": {
      "External id": 533642,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315057.215, "dur": 23.712,
    "args": {
      "External id": 533643,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622315085.810, "dur": 8.975,
    "args": {
      "External id": 533644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622315087.979, "dur": 5.799,
    "args": {
      "External id": 533645,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315090.233, "dur": 3.318,
    "args": {
      "External id": 533646,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622315112.425, "dur": 52.334,
    "args": {
      "External id": 533647,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622315114.339, "dur": 49.366,
    "args": {
      "External id": 533648,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315169.514, "dur": 15.827,
    "args": {
      "External id": 533649,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622315191.807, "dur": 5.635,
    "args": {
      "External id": 533650,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315195.145, "dur": 1.140,
    "args": {
      "External id": 533651,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622315201.675, "dur": 50.557,
    "args": {
      "External id": 533652,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622315202.841, "dur": 4.445,
    "args": {
      "External id": 533653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622315203.988, "dur": 2.702,
    "args": {
      "External id": 533654,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315205.961, "dur": 0.600,
    "args": {
      "External id": 533655,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622315207.967, "dur": 43.858,
    "args": {
      "External id": 533656,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622315209.098, "dur": 42.193,
    "args": {
      "External id": 533657,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622315256.372, "dur": 4.046,
    "args": {
      "External id": 533658,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315258.526, "dur": 0.562,
    "args": {
      "External id": 533659,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622315266.674, "dur": 1.672,
    "args": {
      "External id": 533660,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622315276.007, "dur": 7.545,
    "args": {
      "External id": 533661,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622315278.714, "dur": 4.495,
    "args": {
      "External id": 533662,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622315381.224, "dur": 185.202,
    "args": {
      "External id": 533663,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622315383.364, "dur": 2.275,
    "args": {
      "External id": 533664,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622315389.327, "dur": 176.527,
    "args": {
      "External id": 533665,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622315390.607, "dur": 0.344,
    "args": {
      "External id": 533666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622315392.030, "dur": 20.567,
    "args": {
      "External id": 533667,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622315414.483, "dur": 5.586,
    "args": {
      "External id": 533668,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315416.527, "dur": 3.273,
    "args": {
      "External id": 533669,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622315420.918, "dur": 31.864,
    "args": {
      "External id": 533670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622315421.931, "dur": 2.029,
    "args": {
      "External id": 533671,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622315435.124, "dur": 17.341,
    "args": {
      "External id": 533672,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315437.912, "dur": 2.873,
    "args": {
      "External id": 533673,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622315454.384, "dur": 19.222,
    "args": {
      "External id": 533674,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315475.154, "dur": 13.880,
    "args": {
      "External id": 533675,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622315491.385, "dur": 12.284,
    "args": {
      "External id": 533676,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315505.234, "dur": 11.730,
    "args": {
      "External id": 533677,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622315518.583, "dur": 22.324,
    "args": {
      "External id": 533678,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315523.092, "dur": 1.561,
    "args": {
      "External id": 533679,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315527.280, "dur": 0.809,
    "args": {
      "External id": 533680,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315542.491, "dur": 10.953,
    "args": {
      "External id": 533681,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315554.489, "dur": 10.475,
    "args": {
      "External id": 533682,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622315572.799, "dur": 1.675,
    "args": {
      "External id": 533683,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622315582.649, "dur": 3.617,
    "args": {
      "External id": 533684,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315584.957, "dur": 0.437,
    "args": {
      "External id": 533685,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622315690.755, "dur": 49.965,
    "args": {
      "External id": 533686,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622315746.312, "dur": 6.029,
    "args": {
      "External id": 533687,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315749.455, "dur": 1.306,
    "args": {
      "External id": 533688,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315753.723, "dur": 25.803,
    "args": {
      "External id": 533689,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622315785.115, "dur": 6.659,
    "args": {
      "External id": 533690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622315786.760, "dur": 4.216,
    "args": {
      "External id": 533691,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315788.983, "dur": 1.771,
    "args": {
      "External id": 533692,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622315794.410, "dur": 55.997,
    "args": {
      "External id": 533693,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622315795.272, "dur": 54.627,
    "args": {
      "External id": 533694,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622315854.048, "dur": 15.622,
    "args": {
      "External id": 533695,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622315875.970, "dur": 4.322,
    "args": {
      "External id": 533696,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315878.283, "dur": 1.010,
    "args": {
      "External id": 533697,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622315884.120, "dur": 47.317,
    "args": {
      "External id": 533698,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622315885.224, "dur": 3.521,
    "args": {
      "External id": 533699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622315885.861, "dur": 2.363,
    "args": {
      "External id": 533700,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315887.521, "dur": 0.557,
    "args": {
      "External id": 533701,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622315889.391, "dur": 41.647,
    "args": {
      "External id": 533702,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622315890.027, "dur": 40.514,
    "args": {
      "External id": 533703,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622315935.783, "dur": 3.523,
    "args": {
      "External id": 533704,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622315937.516, "dur": 0.724,
    "args": {
      "External id": 533705,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622315945.494, "dur": 1.468,
    "args": {
      "External id": 533706,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622315955.080, "dur": 6.487,
    "args": {
      "External id": 533707,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622315957.650, "dur": 3.650,
    "args": {
      "External id": 533708,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622316049.714, "dur": 196.711,
    "args": {
      "External id": 533709,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622316053.907, "dur": 2.058,
    "args": {
      "External id": 533710,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622316057.534, "dur": 188.351,
    "args": {
      "External id": 533711,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622316058.588, "dur": 0.541,
    "args": {
      "External id": 533712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622316060.613, "dur": 20.714,
    "args": {
      "External id": 533713,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622316082.849, "dur": 5.361,
    "args": {
      "External id": 533714,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316086.706, "dur": 1.260,
    "args": {
      "External id": 533715,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622316089.013, "dur": 38.525,
    "args": {
      "External id": 533716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622316090.285, "dur": 1.446,
    "args": {
      "External id": 533717,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622316092.873, "dur": 34.257,
    "args": {
      "External id": 533718,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316095.592, "dur": 18.074,
    "args": {
      "External id": 533719,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622316129.745, "dur": 22.073,
    "args": {
      "External id": 533720,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316153.359, "dur": 14.132,
    "args": {
      "External id": 533721,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622316170.159, "dur": 14.066,
    "args": {
      "External id": 533722,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316185.590, "dur": 11.326,
    "args": {
      "External id": 533723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622316199.363, "dur": 20.504,
    "args": {
      "External id": 533724,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316201.662, "dur": 1.766,
    "args": {
      "External id": 533725,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316205.855, "dur": 1.137,
    "args": {
      "External id": 533726,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316221.906, "dur": 10.943,
    "args": {
      "External id": 533727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316233.980, "dur": 10.359,
    "args": {
      "External id": 533728,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622316253.449, "dur": 2.184,
    "args": {
      "External id": 533729,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622316265.203, "dur": 3.119,
    "args": {
      "External id": 533730,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316267.244, "dur": 0.318,
    "args": {
      "External id": 533731,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622316335.297, "dur": 45.736,
    "args": {
      "External id": 533732,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622316385.062, "dur": 5.457,
    "args": {
      "External id": 533733,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316388.179, "dur": 1.142,
    "args": {
      "External id": 533734,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316391.780, "dur": 22.186,
    "args": {
      "External id": 533735,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622316418.219, "dur": 6.305,
    "args": {
      "External id": 533736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622316419.663, "dur": 4.068,
    "args": {
      "External id": 533737,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316422.004, "dur": 1.505,
    "args": {
      "External id": 533738,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622316427.029, "dur": 41.791,
    "args": {
      "External id": 533739,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622316428.398, "dur": 39.739,
    "args": {
      "External id": 533740,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316472.397, "dur": 13.981,
    "args": {
      "External id": 533741,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622316491.472, "dur": 4.595,
    "args": {
      "External id": 533742,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316494.046, "dur": 0.982,
    "args": {
      "External id": 533743,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622316499.572, "dur": 46.825,
    "args": {
      "External id": 533744,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622316500.386, "dur": 3.877,
    "args": {
      "External id": 533745,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622316501.325, "dur": 2.370,
    "args": {
      "External id": 533746,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316502.858, "dur": 0.658,
    "args": {
      "External id": 533747,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622316504.975, "dur": 41.009,
    "args": {
      "External id": 533748,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622316505.881, "dur": 39.481,
    "args": {
      "External id": 533749,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622316549.822, "dur": 4.086,
    "args": {
      "External id": 533750,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316551.775, "dur": 1.015,
    "args": {
      "External id": 533751,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622316559.584, "dur": 1.515,
    "args": {
      "External id": 533752,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622316568.350, "dur": 9.391,
    "args": {
      "External id": 533753,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622316571.196, "dur": 6.227,
    "args": {
      "External id": 533754,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622316707.508, "dur": 185.185,
    "args": {
      "External id": 533755,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622316710.136, "dur": 3.889,
    "args": {
      "External id": 533756,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622316716.209, "dur": 175.885,
    "args": {
      "External id": 533757,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622316717.899, "dur": 0.269,
    "args": {
      "External id": 533758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622316719.743, "dur": 25.278,
    "args": {
      "External id": 533759,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622316746.575, "dur": 4.159,
    "args": {
      "External id": 533760,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316748.916, "dur": 1.527,
    "args": {
      "External id": 533761,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622316751.496, "dur": 24.216,
    "args": {
      "External id": 533762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622316753.083, "dur": 1.667,
    "args": {
      "External id": 533763,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622316755.878, "dur": 19.492,
    "args": {
      "External id": 533764,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316761.028, "dur": 2.666,
    "args": {
      "External id": 533765,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622316777.303, "dur": 20.119,
    "args": {
      "External id": 533766,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316799.032, "dur": 13.238,
    "args": {
      "External id": 533767,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622316815.067, "dur": 13.288,
    "args": {
      "External id": 533768,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316830.117, "dur": 11.944,
    "args": {
      "External id": 533769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622316843.841, "dur": 19.160,
    "args": {
      "External id": 533770,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316845.706, "dur": 1.441,
    "args": {
      "External id": 533771,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316849.529, "dur": 1.084,
    "args": {
      "External id": 533772,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316865.193, "dur": 13.345,
    "args": {
      "External id": 533773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622316879.760, "dur": 11.011,
    "args": {
      "External id": 533774,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622316899.513, "dur": 2.249,
    "args": {
      "External id": 533775,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622316910.896, "dur": 3.541,
    "args": {
      "External id": 533776,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622316913.151, "dur": 0.483,
    "args": {
      "External id": 533777,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622316985.601, "dur": 47.456,
    "args": {
      "External id": 533778,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622317037.327, "dur": 5.588,
    "args": {
      "External id": 533779,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317040.493, "dur": 1.392,
    "args": {
      "External id": 533780,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317044.224, "dur": 24.031,
    "args": {
      "External id": 533781,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622317073.260, "dur": 6.149,
    "args": {
      "External id": 533782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622317075.006, "dur": 3.765,
    "args": {
      "External id": 533783,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317077.564, "dur": 0.988,
    "args": {
      "External id": 533784,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622317082.047, "dur": 64.458,
    "args": {
      "External id": 533785,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622317083.158, "dur": 62.249,
    "args": {
      "External id": 533786,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317151.677, "dur": 16.541,
    "args": {
      "External id": 533787,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622317174.570, "dur": 5.035,
    "args": {
      "External id": 533788,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317177.389, "dur": 1.004,
    "args": {
      "External id": 533789,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622317183.519, "dur": 67.018,
    "args": {
      "External id": 533790,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622317184.563, "dur": 6.525,
    "args": {
      "External id": 533791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622317185.840, "dur": 4.630,
    "args": {
      "External id": 533792,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317187.621, "dur": 2.654,
    "args": {
      "External id": 533793,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622317205.752, "dur": 44.341,
    "args": {
      "External id": 533794,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622317206.499, "dur": 43.076,
    "args": {
      "External id": 533795,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622317254.930, "dur": 3.703,
    "args": {
      "External id": 533796,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317256.815, "dur": 0.803,
    "args": {
      "External id": 533797,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622317265.337, "dur": 1.564,
    "args": {
      "External id": 533798,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 7957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622317274.842, "dur": 6.100,
    "args": {
      "External id": 533799,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 7958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622317276.849, "dur": 3.818,
    "args": {
      "External id": 533800,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622317368.421, "dur": 191.230,
    "args": {
      "External id": 533801,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622317372.698, "dur": 2.376,
    "args": {
      "External id": 533802,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622317376.663, "dur": 182.417,
    "args": {
      "External id": 533803,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 7962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622317378.050, "dur": 0.375,
    "args": {
      "External id": 533804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622317379.835, "dur": 23.991,
    "args": {
      "External id": 533805,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 7964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622317405.524, "dur": 5.307,
    "args": {
      "External id": 533806,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 7965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317409.615, "dur": 0.912,
    "args": {
      "External id": 533807,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 7966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622317411.801, "dur": 22.867,
    "args": {
      "External id": 533808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622317412.898, "dur": 1.840,
    "args": {
      "External id": 533809,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622317416.092, "dur": 18.213,
    "args": {
      "External id": 533810,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 7969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317418.692, "dur": 2.613,
    "args": {
      "External id": 533811,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622317436.348, "dur": 20.913,
    "args": {
      "External id": 533812,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317458.686, "dur": 15.241,
    "args": {
      "External id": 533813,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622317476.213, "dur": 15.570,
    "args": {
      "External id": 533814,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 7973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317493.140, "dur": 12.826,
    "args": {
      "External id": 533815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 7974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622317507.778, "dur": 22.418,
    "args": {
      "External id": 533816,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 7975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317509.447, "dur": 1.213,
    "args": {
      "External id": 533817,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 7976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317512.991, "dur": 0.937,
    "args": {
      "External id": 533818,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 7977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317531.870, "dur": 12.671,
    "args": {
      "External id": 533819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 7978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317545.850, "dur": 12.122,
    "args": {
      "External id": 533820,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 7979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622317565.321, "dur": 1.530,
    "args": {
      "External id": 533821,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 7980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622317574.946, "dur": 3.684,
    "args": {
      "External id": 533822,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 7981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317577.343, "dur": 0.434,
    "args": {
      "External id": 533823,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 7982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622317638.286, "dur": 85.612,
    "args": {
      "External id": 533824,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 7983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622317731.102, "dur": 6.496,
    "args": {
      "External id": 533825,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317734.576, "dur": 1.208,
    "args": {
      "External id": 533826,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317738.944, "dur": 25.948,
    "args": {
      "External id": 533827,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 7986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622317770.624, "dur": 6.299,
    "args": {
      "External id": 533828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 7987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622317772.084, "dur": 4.194,
    "args": {
      "External id": 533829,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 7988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317774.355, "dur": 1.731,
    "args": {
      "External id": 533830,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 7989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622317779.954, "dur": 44.083,
    "args": {
      "External id": 533831,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622317781.056, "dur": 42.270,
    "args": {
      "External id": 533832,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 7991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622317827.793, "dur": 16.565,
    "args": {
      "External id": 533833,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 7992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622317850.267, "dur": 4.290,
    "args": {
      "External id": 533834,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 7993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317852.686, "dur": 0.819,
    "args": {
      "External id": 533835,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 7994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622317858.323, "dur": 48.932,
    "args": {
      "External id": 533836,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 7995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622317859.569, "dur": 4.403,
    "args": {
      "External id": 533837,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 7996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622317860.272, "dur": 3.130,
    "args": {
      "External id": 533838,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 7997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317861.902, "dur": 1.185,
    "args": {
      "External id": 533839,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 7998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622317864.714, "dur": 42.090,
    "args": {
      "External id": 533840,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 7999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622317865.226, "dur": 40.955,
    "args": {
      "External id": 533841,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622317911.010, "dur": 3.727,
    "args": {
      "External id": 533842,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622317912.827, "dur": 0.849,
    "args": {
      "External id": 533843,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622317921.024, "dur": 1.590,
    "args": {
      "External id": 533844,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622317930.193, "dur": 9.057,
    "args": {
      "External id": 533845,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622317932.276, "dur": 6.645,
    "args": {
      "External id": 533846,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622318024.272, "dur": 199.422,
    "args": {
      "External id": 533847,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622318026.476, "dur": 2.561,
    "args": {
      "External id": 533848,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622318030.839, "dur": 192.312,
    "args": {
      "External id": 533849,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622318032.572, "dur": 0.295,
    "args": {
      "External id": 533850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622318034.226, "dur": 23.552,
    "args": {
      "External id": 533851,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622318059.060, "dur": 2.998,
    "args": {
      "External id": 533852,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318060.990, "dur": 0.686,
    "args": {
      "External id": 533853,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622318062.976, "dur": 22.347,
    "args": {
      "External id": 533854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622318064.265, "dur": 1.884,
    "args": {
      "External id": 533855,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622318067.216, "dur": 17.754,
    "args": {
      "External id": 533856,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318072.085, "dur": 2.331,
    "args": {
      "External id": 533857,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622318087.283, "dur": 39.654,
    "args": {
      "External id": 533858,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318129.229, "dur": 14.019,
    "args": {
      "External id": 533859,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622318146.349, "dur": 14.016,
    "args": {
      "External id": 533860,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318161.989, "dur": 11.640,
    "args": {
      "External id": 533861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622318175.673, "dur": 19.571,
    "args": {
      "External id": 533862,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318177.403, "dur": 1.791,
    "args": {
      "External id": 533863,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318181.451, "dur": 0.837,
    "args": {
      "External id": 533864,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318196.848, "dur": 13.141,
    "args": {
      "External id": 533865,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318211.288, "dur": 10.870,
    "args": {
      "External id": 533866,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622318230.392, "dur": 2.192,
    "args": {
      "External id": 533867,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622318240.623, "dur": 3.307,
    "args": {
      "External id": 533868,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318242.594, "dur": 0.475,
    "args": {
      "External id": 533869,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622318312.373, "dur": 44.346,
    "args": {
      "External id": 533870,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622318361.346, "dur": 5.301,
    "args": {
      "External id": 533871,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318364.358, "dur": 1.271,
    "args": {
      "External id": 533872,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318367.883, "dur": 22.477,
    "args": {
      "External id": 533873,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622318394.924, "dur": 5.606,
    "args": {
      "External id": 533874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622318396.351, "dur": 3.603,
    "args": {
      "External id": 533875,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318398.630, "dur": 1.130,
    "args": {
      "External id": 533876,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622318403.068, "dur": 40.281,
    "args": {
      "External id": 533877,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622318404.077, "dur": 38.658,
    "args": {
      "External id": 533878,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318446.897, "dur": 15.597,
    "args": {
      "External id": 533879,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622318467.995, "dur": 4.462,
    "args": {
      "External id": 533880,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318470.652, "dur": 0.836,
    "args": {
      "External id": 533881,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622318475.781, "dur": 49.449,
    "args": {
      "External id": 533882,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622318477.058, "dur": 6.313,
    "args": {
      "External id": 533883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622318477.830, "dur": 5.001,
    "args": {
      "External id": 533884,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318479.595, "dur": 3.031,
    "args": {
      "External id": 533885,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622318484.011, "dur": 40.836,
    "args": {
      "External id": 533886,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622318484.493, "dur": 39.836,
    "args": {
      "External id": 533887,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622318528.827, "dur": 3.359,
    "args": {
      "External id": 533888,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318530.443, "dur": 0.658,
    "args": {
      "External id": 533889,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622318537.501, "dur": 1.232,
    "args": {
      "External id": 533890,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622318545.938, "dur": 6.175,
    "args": {
      "External id": 533891,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622318548.188, "dur": 3.652,
    "args": {
      "External id": 533892,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622318629.237, "dur": 262.996,
    "args": {
      "External id": 533893,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622318631.401, "dur": 2.699,
    "args": {
      "External id": 533894,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622318635.971, "dur": 255.827,
    "args": {
      "External id": 533895,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622318637.484, "dur": 0.297,
    "args": {
      "External id": 533896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622318640.824, "dur": 74.452,
    "args": {
      "External id": 533897,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622318718.013, "dur": 3.365,
    "args": {
      "External id": 533898,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318720.123, "dur": 0.942,
    "args": {
      "External id": 533899,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622318722.274, "dur": 27.209,
    "args": {
      "External id": 533900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622318723.561, "dur": 3.988,
    "args": {
      "External id": 533901,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622318728.872, "dur": 20.245,
    "args": {
      "External id": 533902,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318731.981, "dur": 2.638,
    "args": {
      "External id": 533903,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622318751.162, "dur": 25.300,
    "args": {
      "External id": 533904,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318777.839, "dur": 30.647,
    "args": {
      "External id": 533905,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622318810.616, "dur": 14.992,
    "args": {
      "External id": 533906,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318826.852, "dur": 13.394,
    "args": {
      "External id": 533907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622318841.907, "dur": 21.559,
    "args": {
      "External id": 533908,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318844.186, "dur": 1.413,
    "args": {
      "External id": 533909,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318847.716, "dur": 1.170,
    "args": {
      "External id": 533910,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318864.873, "dur": 12.572,
    "args": {
      "External id": 533911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622318878.632, "dur": 11.813,
    "args": {
      "External id": 533912,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622318899.250, "dur": 2.287,
    "args": {
      "External id": 533913,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622318910.351, "dur": 3.865,
    "args": {
      "External id": 533914,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622318912.938, "dur": 0.434,
    "args": {
      "External id": 533915,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622318980.409, "dur": 48.399,
    "args": {
      "External id": 533916,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622319045.780, "dur": 6.760,
    "args": {
      "External id": 533917,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319048.707, "dur": 2.826,
    "args": {
      "External id": 533918,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319053.951, "dur": 31.235,
    "args": {
      "External id": 533919,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622319089.906, "dur": 5.738,
    "args": {
      "External id": 533920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622319091.386, "dur": 3.402,
    "args": {
      "External id": 533921,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319093.284, "dur": 1.242,
    "args": {
      "External id": 533922,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622319113.478, "dur": 49.827,
    "args": {
      "External id": 533923,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622319114.759, "dur": 47.659,
    "args": {
      "External id": 533924,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319167.538, "dur": 16.910,
    "args": {
      "External id": 533925,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622319190.671, "dur": 5.087,
    "args": {
      "External id": 533926,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319193.508, "dur": 0.854,
    "args": {
      "External id": 533927,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622319199.504, "dur": 49.585,
    "args": {
      "External id": 533928,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622319200.631, "dur": 4.645,
    "args": {
      "External id": 533929,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622319201.619, "dur": 3.102,
    "args": {
      "External id": 533930,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319203.449, "dur": 1.121,
    "args": {
      "External id": 533931,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622319205.890, "dur": 42.766,
    "args": {
      "External id": 533932,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622319206.404, "dur": 41.708,
    "args": {
      "External id": 533933,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622319253.322, "dur": 3.462,
    "args": {
      "External id": 533934,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319255.036, "dur": 0.754,
    "args": {
      "External id": 533935,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622319262.695, "dur": 1.494,
    "args": {
      "External id": 533936,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622319271.803, "dur": 7.062,
    "args": {
      "External id": 533937,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622319273.832, "dur": 4.714,
    "args": {
      "External id": 533938,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622319367.653, "dur": 188.809,
    "args": {
      "External id": 533939,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622319369.858, "dur": 4.068,
    "args": {
      "External id": 533940,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622319375.257, "dur": 180.528,
    "args": {
      "External id": 533941,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622319376.733, "dur": 0.523,
    "args": {
      "External id": 533942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622319378.430, "dur": 22.483,
    "args": {
      "External id": 533943,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622319402.677, "dur": 5.322,
    "args": {
      "External id": 533944,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319406.865, "dur": 0.874,
    "args": {
      "External id": 533945,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622319408.983, "dur": 23.102,
    "args": {
      "External id": 533946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622319410.253, "dur": 1.756,
    "args": {
      "External id": 533947,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622319413.447, "dur": 18.317,
    "args": {
      "External id": 533948,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319416.177, "dur": 2.508,
    "args": {
      "External id": 533949,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622319433.791, "dur": 21.046,
    "args": {
      "External id": 533950,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319456.159, "dur": 13.985,
    "args": {
      "External id": 533951,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622319472.710, "dur": 14.457,
    "args": {
      "External id": 533952,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319488.503, "dur": 13.331,
    "args": {
      "External id": 533953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622319503.522, "dur": 23.167,
    "args": {
      "External id": 533954,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319507.778, "dur": 1.413,
    "args": {
      "External id": 533955,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319511.426, "dur": 0.775,
    "args": {
      "External id": 533956,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319528.289, "dur": 12.466,
    "args": {
      "External id": 533957,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319541.942, "dur": 12.953,
    "args": {
      "External id": 533958,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622319562.007, "dur": 1.928,
    "args": {
      "External id": 533959,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622319572.913, "dur": 3.907,
    "args": {
      "External id": 533960,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319575.553, "dur": 0.476,
    "args": {
      "External id": 533961,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622319642.659, "dur": 94.784,
    "args": {
      "External id": 533962,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622319744.560, "dur": 7.148,
    "args": {
      "External id": 533963,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319748.310, "dur": 1.594,
    "args": {
      "External id": 533964,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319753.000, "dur": 23.664,
    "args": {
      "External id": 533965,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622319782.166, "dur": 5.913,
    "args": {
      "External id": 533966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622319783.605, "dur": 3.776,
    "args": {
      "External id": 533967,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319785.798, "dur": 1.398,
    "args": {
      "External id": 533968,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622319790.910, "dur": 45.117,
    "args": {
      "External id": 533969,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622319792.087, "dur": 43.085,
    "args": {
      "External id": 533970,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319839.604, "dur": 16.383,
    "args": {
      "External id": 533971,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622319860.981, "dur": 31.583,
    "args": {
      "External id": 533972,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622319863.411, "dur": 28.786,
    "args": {
      "External id": 533973,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319868.544, "dur": 1.043,
    "args": {
      "External id": 533974,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622319897.625, "dur": 31.952,
    "args": {
      "External id": 533975,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 8134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622319899.413, "dur": 29.951,
    "args": {
      "External id": 533976,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], [], []], "Ev Idx": 8135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622319904.621, "dur": 4.532,
    "args": {
      "External id": 533977,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319910.246, "dur": 18.468,
    "args": {
      "External id": 533978,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622319940.903, "dur": 5.070,
    "args": {
      "External id": 533979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 8138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622319943.086, "dur": 2.622,
    "args": {
      "External id": 533980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 8139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622319947.117, "dur": 1.518,
    "args": {
      "External id": 533981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622319948.061, "dur": 0.484,
    "args": {
      "External id": 533982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622319992.221, "dur": 22.401,
    "args": {
      "External id": 533983,"Sequence number": 6319620, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 8142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320016.770, "dur": 14.632,
    "args": {
      "External id": 533984,"Sequence number": 6319621, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 8143
    }
  },
  {
    "ph": "s", "id": 9, "pid": 4183438, "tid": 4183438, "ts": 679622320016.770,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622320037.012, "dur": 5.677,
    "args": {
      "External id": 533985,"Sequence number": 6319622, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 8144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320040.450, "dur": 0.864,
    "args": {
      "External id": 533986,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 8145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622320045.514, "dur": 6.676,
    "args": {
      "External id": 533987,"Sequence number": 6319622, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "3"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 8146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320049.992, "dur": 0.917,
    "args": {
      "External id": 533988,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "4"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 8147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622320053.456, "dur": 3.187,
    "args": {
      "External id": 533989,"Sequence number": 6319622, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 8148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320055.383, "dur": 0.664,
    "args": {
      "External id": 533990,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "4"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 8149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622320060.849, "dur": 6.563,
    "args": {
      "External id": 533991,"Sequence number": 6319622, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 8150
    }
  },
  {
    "ph": "s", "id": 8, "pid": 4183438, "tid": 4183438, "ts": 679622320060.849,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320065.063, "dur": 1.060,
    "args": {
      "External id": 533992,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 8151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622320068.349, "dur": 5.364,
    "args": {
      "External id": 533993,"Sequence number": 6319623, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], [], []], "Ev Idx": 8152
    }
  },
  {
    "ph": "s", "id": 7, "pid": 4183438, "tid": 4183438, "ts": 679622320068.349,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320072.231, "dur": 0.738,
    "args": {
      "External id": 533994,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 1024]", "[16777216, 4096, 1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 8153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622320074.723, "dur": 5.075,
    "args": {
      "External id": 533995,"Sequence number": 6319624, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "3"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], []], "Input Dims": [[16, 4096, 4, 1024], [], []], "Ev Idx": 8154
    }
  },
  {
    "ph": "s", "id": 6, "pid": 4183438, "tid": 4183438, "ts": 679622320074.723,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320078.215, "dur": 0.770,
    "args": {
      "External id": 533996,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "3072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1024, 1], [], [], []], "Input Dims": [[16, 4096, 4, 1024], [], [], []], "Ev Idx": 8155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622320080.889, "dur": 4.543,
    "args": {
      "External id": 533997,"Sequence number": 6319625, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], [], []], "Ev Idx": 8156
    }
  },
  {
    "ph": "s", "id": 5, "pid": 4183438, "tid": 4183438, "ts": 679622320080.889,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320083.685, "dur": 1.012,
    "args": {
      "External id": 533998,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1024]", "[16777216, 4096, 1]", "3072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[16777216, 4096, 1], [], [], []], "Input Dims": [[16, 4096, 1024], [], [], []], "Ev Idx": 8157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622320089.002, "dur": 51.644,
    "args": {
      "External id": 533999,"Sequence number": 6319626, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 8158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622320090.692, "dur": 49.687,
    "args": {
      "External id": 534000,"Sequence number": 6319626, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 8159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622320092.928, "dur": 25.124,
    "args": {
      "External id": 534001,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 8160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622320095.785, "dur": 21.243,
    "args": {
      "External id": 534002,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320119.511, "dur": 20.308,
    "args": {
      "External id": 534003,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 8162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622320168.415, "dur": 5.352,
    "args": {
      "External id": 534004,"Sequence number": 6319626, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[16777216, 4096, 1], []], "Input Dims": [[16, 4096, 1024], []], "Ev Idx": 8163
    }
  },
  {
    "ph": "s", "id": 4, "pid": 4183438, "tid": 4183438, "ts": 679622320168.415,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622320176.446, "dur": 1.219,
    "args": {
      "External id": 534005,"Sequence number": 6319627, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 8164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 4183438, "tid": 4183438,
    "ts": 679622320209.880, "dur": 24114.709,
    "args": {
      "External id": 534006,"Sequence number": 6319627, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1], [1], [1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [65536], [32000, 1024], [], [], [], [], []], "Ev Idx": 8165
    }
  },
  {
    "ph": "s", "id": 3, "pid": 4183438, "tid": 4183438, "ts": 679622320209.880,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 4183438, "tid": 4183438,
    "ts": 679622320223.994, "dur": 27.760,
    "args": {
      "External id": 534007,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 8166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622320224.911, "dur": 26.656,
    "args": {
      "External id": 534008,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 8167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622320226.859, "dur": 5.775,
    "args": {
      "External id": 534009,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[4096, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 8168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622320228.453, "dur": 3.779,
    "args": {
      "External id": 534010,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320233.500, "dur": 17.634,
    "args": {
      "External id": 534011,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [4096, 1], []], "Input Dims": [[65536, 1024], [65536, 1024], []], "Ev Idx": 8170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622320267.725, "dur": 27.190,
    "args": {
      "External id": 534012,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 8171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622320268.878, "dur": 6.326,
    "args": {
      "External id": 534013,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], [], []], "Ev Idx": 8172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320271.119, "dur": 3.772,
    "args": {
      "External id": 534014,"Record function id": 0, "Concrete Inputs": ["[65536, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320276.680, "dur": 18.024,
    "args": {
      "External id": 534015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 8174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320278.462, "dur": 15.853,
    "args": {
      "External id": 534016,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[65536, 1024], []], "Ev Idx": 8175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622320323.585, "dur": 25.913,
    "args": {
      "External id": 534017,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 8176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622320324.383, "dur": 4.713,
    "args": {
      "External id": 534018,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 8177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320326.045, "dur": 2.831,
    "args": {
      "External id": 534019,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320329.748, "dur": 19.543,
    "args": {
      "External id": 534020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320330.406, "dur": 18.423,
    "args": {
      "External id": 534021,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1024, 1], []], "Input Dims": [[32000, 1024], []], "Ev Idx": 8180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622320355.480, "dur": 17.878,
    "args": {
      "External id": 534022,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 8181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622320357.301, "dur": 2.641,
    "args": {
      "External id": 534023,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320360.499, "dur": 12.578,
    "args": {
      "External id": 534024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 8183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320361.069, "dur": 11.698,
    "args": {
      "External id": 534025,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 4183438,
    "ts": 679622320378.118, "dur": 22.065,
    "args": {
      "External id": 534026,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622320403.075, "dur": 48.697,
    "args": {
      "External id": 534027,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622320405.155, "dur": 46.110,
    "args": {
      "External id": 534028,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320408.998, "dur": 1.073,
    "args": {
      "External id": 534029,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622320411.391, "dur": 24.098,
    "args": {
      "External id": 534030,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622320415.744, "dur": 19.511,
    "args": {
      "External id": 534031,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 8190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622320417.960, "dur": 2.545,
    "args": {
      "External id": 534032,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622320421.725, "dur": 13.190,
    "args": {
      "External id": 534033,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 8192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622320455.318, "dur": 18097.309,
    "args": {
      "External id": 534034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 8193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622320456.848, "dur": 18094.072,
    "args": {
      "External id": 534035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 8194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622338567.391, "dur": 9.373,
    "args": {
      "External id": 534036,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622338573.330, "dur": 1.302,
    "args": {
      "External id": 534037,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622338582.689, "dur": 133.596,
    "args": {
      "External id": 534038,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622338584.250, "dur": 6.639,
    "args": {
      "External id": 534039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622338586.583, "dur": 3.452,
    "args": {
      "External id": 534040,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622338588.595, "dur": 1.148,
    "args": {
      "External id": 534041,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622338592.329, "dur": 122.908,
    "args": {
      "External id": 534042,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622338594.590, "dur": 119.163,
    "args": {
      "External id": 534043,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622338722.172, "dur": 5.461,
    "args": {
      "External id": 534044,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622338724.927, "dur": 0.939,
    "args": {
      "External id": 534045,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622338736.811, "dur": 3.100,
    "args": {
      "External id": 534046,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622338754.644, "dur": 7.923,
    "args": {
      "External id": 534047,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622338756.979, "dur": 5.269,
    "args": {
      "External id": 534048,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622338896.313, "dur": 230.279,
    "args": {
      "External id": 534049,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622338901.312, "dur": 2.730,
    "args": {
      "External id": 534050,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622338905.767, "dur": 220.240,
    "args": {
      "External id": 534051,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622338907.641, "dur": 0.461,
    "args": {
      "External id": 534052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622338909.829, "dur": 29.257,
    "args": {
      "External id": 534053,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622338941.028, "dur": 5.985,
    "args": {
      "External id": 534054,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622338945.649, "dur": 1.044,
    "args": {
      "External id": 534055,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622338948.358, "dur": 25.654,
    "args": {
      "External id": 534056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622338949.941, "dur": 1.412,
    "args": {
      "External id": 534057,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622338952.779, "dur": 20.921,
    "args": {
      "External id": 534058,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622338956.666, "dur": 2.990,
    "args": {
      "External id": 534059,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622338975.983, "dur": 25.856,
    "args": {
      "External id": 534060,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339003.793, "dur": 16.359,
    "args": {
      "External id": 534061,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622339023.527, "dur": 16.522,
    "args": {
      "External id": 534062,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339041.742, "dur": 14.021,
    "args": {
      "External id": 534063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622339057.952, "dur": 23.358,
    "args": {
      "External id": 534064,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339060.237, "dur": 1.842,
    "args": {
      "External id": 534065,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339065.192, "dur": 1.162,
    "args": {
      "External id": 534066,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339082.920, "dur": 13.071,
    "args": {
      "External id": 534067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339109.650, "dur": 14.704,
    "args": {
      "External id": 534068,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622339134.880, "dur": 2.564,
    "args": {
      "External id": 534069,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622339144.750, "dur": 4.406,
    "args": {
      "External id": 534070,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339147.345, "dur": 0.865,
    "args": {
      "External id": 534071,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622339230.656, "dur": 60.194,
    "args": {
      "External id": 534072,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622339296.116, "dur": 13.409,
    "args": {
      "External id": 534073,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339305.246, "dur": 1.323,
    "args": {
      "External id": 534074,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339311.519, "dur": 26.537,
    "args": {
      "External id": 534075,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622339344.161, "dur": 8.792,
    "args": {
      "External id": 534076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622339346.025, "dur": 6.111,
    "args": {
      "External id": 534077,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339348.714, "dur": 3.125,
    "args": {
      "External id": 534078,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622339356.072, "dur": 48.285,
    "args": {
      "External id": 534079,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622339357.131, "dur": 46.526,
    "args": {
      "External id": 534080,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339408.923, "dur": 17.752,
    "args": {
      "External id": 534081,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622339432.803, "dur": 4.677,
    "args": {
      "External id": 534082,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339435.702, "dur": 0.865,
    "args": {
      "External id": 534083,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622339441.851, "dur": 48.707,
    "args": {
      "External id": 534084,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622339443.010, "dur": 4.272,
    "args": {
      "External id": 534085,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622339443.867, "dur": 2.837,
    "args": {
      "External id": 534086,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339445.707, "dur": 0.824,
    "args": {
      "External id": 534087,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622339447.945, "dur": 42.253,
    "args": {
      "External id": 534088,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622339448.502, "dur": 41.229,
    "args": {
      "External id": 534089,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622339494.765, "dur": 3.797,
    "args": {
      "External id": 534090,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339496.703, "dur": 0.585,
    "args": {
      "External id": 534091,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622339505.112, "dur": 1.628,
    "args": {
      "External id": 534092,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622339514.913, "dur": 7.404,
    "args": {
      "External id": 534093,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622339517.176, "dur": 4.854,
    "args": {
      "External id": 534094,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622339617.127, "dur": 319.659,
    "args": {
      "External id": 534095,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622339619.837, "dur": 2.190,
    "args": {
      "External id": 534096,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622339625.744, "dur": 310.334,
    "args": {
      "External id": 534097,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622339627.422, "dur": 0.499,
    "args": {
      "External id": 534098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622339629.501, "dur": 58.598,
    "args": {
      "External id": 534099,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622339691.254, "dur": 6.159,
    "args": {
      "External id": 534100,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339693.511, "dur": 3.482,
    "args": {
      "External id": 534101,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622339698.654, "dur": 27.057,
    "args": {
      "External id": 534102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622339700.151, "dur": 2.345,
    "args": {
      "External id": 534103,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622339704.023, "dur": 21.357,
    "args": {
      "External id": 534104,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339707.504, "dur": 3.132,
    "args": {
      "External id": 534105,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622339727.656, "dur": 22.647,
    "args": {
      "External id": 534106,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339751.885, "dur": 13.060,
    "args": {
      "External id": 534107,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622339767.795, "dur": 15.595,
    "args": {
      "External id": 534108,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339784.904, "dur": 19.965,
    "args": {
      "External id": 534109,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622339806.959, "dur": 37.896,
    "args": {
      "External id": 534110,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339808.854, "dur": 1.841,
    "args": {
      "External id": 534111,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339813.191, "dur": 2.724,
    "args": {
      "External id": 534112,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339846.597, "dur": 43.882,
    "args": {
      "External id": 534113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622339891.832, "dur": 43.252,
    "args": {
      "External id": 534114,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622339944.601, "dur": 2.789,
    "args": {
      "External id": 534115,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622339956.873, "dur": 4.224,
    "args": {
      "External id": 534116,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622339959.454, "dur": 0.485,
    "args": {
      "External id": 534117,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622340034.491, "dur": 51.551,
    "args": {
      "External id": 534118,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622340091.092, "dur": 4.906,
    "args": {
      "External id": 534119,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340094.291, "dur": 0.650,
    "args": {
      "External id": 534120,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "8388608"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340112.248, "dur": 29.589,
    "args": {
      "External id": 534121,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622340158.355, "dur": 9.057,
    "args": {
      "External id": 534122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622340160.697, "dur": 5.705,
    "args": {
      "External id": 534123,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340163.393, "dur": 2.785,
    "args": {
      "External id": 534124,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622340170.448, "dur": 47.928,
    "args": {
      "External id": 534125,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622340171.651, "dur": 45.941,
    "args": {
      "External id": 534126,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340223.096, "dur": 16.552,
    "args": {
      "External id": 534127,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622340245.709, "dur": 4.394,
    "args": {
      "External id": 534128,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340248.230, "dur": 1.024,
    "args": {
      "External id": 534129,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622340254.244, "dur": 50.159,
    "args": {
      "External id": 534130,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622340255.281, "dur": 4.488,
    "args": {
      "External id": 534131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622340256.309, "dur": 2.885,
    "args": {
      "External id": 534132,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340258.153, "dur": 0.899,
    "args": {
      "External id": 534133,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622340260.489, "dur": 43.519,
    "args": {
      "External id": 534134,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622340261.417, "dur": 42.046,
    "args": {
      "External id": 534135,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622340308.591, "dur": 3.710,
    "args": {
      "External id": 534136,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340310.411, "dur": 0.804,
    "args": {
      "External id": 534137,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622340318.774, "dur": 1.782,
    "args": {
      "External id": 534138,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622340328.600, "dur": 7.079,
    "args": {
      "External id": 534139,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622340331.154, "dur": 4.245,
    "args": {
      "External id": 534140,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622340424.588, "dur": 191.406,
    "args": {
      "External id": 534141,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622340427.252, "dur": 2.093,
    "args": {
      "External id": 534142,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622340432.932, "dur": 182.620,
    "args": {
      "External id": 534143,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622340434.557, "dur": 0.490,
    "args": {
      "External id": 534144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622340436.323, "dur": 20.439,
    "args": {
      "External id": 534145,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622340458.521, "dur": 5.317,
    "args": {
      "External id": 534146,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340461.080, "dur": 2.499,
    "args": {
      "External id": 534147,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622340464.795, "dur": 23.823,
    "args": {
      "External id": 534148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622340466.096, "dur": 1.894,
    "args": {
      "External id": 534149,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622340469.124, "dur": 19.206,
    "args": {
      "External id": 534150,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340471.720, "dur": 2.571,
    "args": {
      "External id": 534151,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622340493.130, "dur": 20.495,
    "args": {
      "External id": 534152,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340514.970, "dur": 17.874,
    "args": {
      "External id": 534153,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622340535.122, "dur": 15.147,
    "args": {
      "External id": 534154,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340551.617, "dur": 13.058,
    "args": {
      "External id": 534155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622340566.382, "dur": 21.138,
    "args": {
      "External id": 534156,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340568.340, "dur": 1.936,
    "args": {
      "External id": 534157,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340572.327, "dur": 0.808,
    "args": {
      "External id": 534158,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340589.212, "dur": 12.841,
    "args": {
      "External id": 534159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340603.065, "dur": 11.190,
    "args": {
      "External id": 534160,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622340621.935, "dur": 2.143,
    "args": {
      "External id": 534161,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622340632.607, "dur": 3.380,
    "args": {
      "External id": 534162,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340634.749, "dur": 0.539,
    "args": {
      "External id": 534163,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622340742.951, "dur": 54.924,
    "args": {
      "External id": 534164,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622340803.443, "dur": 5.960,
    "args": {
      "External id": 534165,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340806.760, "dur": 1.135,
    "args": {
      "External id": 534166,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340810.849, "dur": 24.837,
    "args": {
      "External id": 534167,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622340841.076, "dur": 6.382,
    "args": {
      "External id": 534168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622340842.862, "dur": 3.992,
    "args": {
      "External id": 534169,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340845.188, "dur": 1.467,
    "args": {
      "External id": 534170,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622340849.955, "dur": 44.079,
    "args": {
      "External id": 534171,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622340851.159, "dur": 42.390,
    "args": {
      "External id": 534172,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622340898.306, "dur": 16.688,
    "args": {
      "External id": 534173,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622340920.390, "dur": 4.304,
    "args": {
      "External id": 534174,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340922.877, "dur": 0.997,
    "args": {
      "External id": 534175,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622340928.537, "dur": 47.671,
    "args": {
      "External id": 534176,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622340929.480, "dur": 3.660,
    "args": {
      "External id": 534177,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622340930.486, "dur": 2.082,
    "args": {
      "External id": 534178,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340931.830, "dur": 0.584,
    "args": {
      "External id": 534179,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622340933.840, "dur": 41.923,
    "args": {
      "External id": 534180,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622340934.877, "dur": 40.475,
    "args": {
      "External id": 534181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622340980.262, "dur": 3.840,
    "args": {
      "External id": 534182,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622340982.255, "dur": 0.702,
    "args": {
      "External id": 534183,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622340990.111, "dur": 1.688,
    "args": {
      "External id": 534184,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622340999.316, "dur": 9.203,
    "args": {
      "External id": 534185,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622341001.764, "dur": 6.390,
    "args": {
      "External id": 534186,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622341089.921, "dur": 209.079,
    "args": {
      "External id": 534187,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622341091.979, "dur": 2.811,
    "args": {
      "External id": 534188,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622341096.563, "dur": 201.834,
    "args": {
      "External id": 534189,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622341112.300, "dur": 0.505,
    "args": {
      "External id": 534190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622341114.641, "dur": 28.556,
    "args": {
      "External id": 534191,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622341145.044, "dur": 3.828,
    "args": {
      "External id": 534192,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341147.310, "dur": 1.271,
    "args": {
      "External id": 534193,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622341149.952, "dur": 25.555,
    "args": {
      "External id": 534194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622341151.081, "dur": 1.823,
    "args": {
      "External id": 534195,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622341154.262, "dur": 20.850,
    "args": {
      "External id": 534196,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341159.840, "dur": 2.637,
    "args": {
      "External id": 534197,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622341177.218, "dur": 22.926,
    "args": {
      "External id": 534198,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341201.444, "dur": 13.058,
    "args": {
      "External id": 534199,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622341217.685, "dur": 14.951,
    "args": {
      "External id": 534200,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341234.366, "dur": 13.521,
    "args": {
      "External id": 534201,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622341249.472, "dur": 20.553,
    "args": {
      "External id": 534202,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341251.447, "dur": 1.545,
    "args": {
      "External id": 534203,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341254.967, "dur": 0.877,
    "args": {
      "External id": 534204,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341271.784, "dur": 13.112,
    "args": {
      "External id": 534205,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341285.907, "dur": 11.133,
    "args": {
      "External id": 534206,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622341306.236, "dur": 2.369,
    "args": {
      "External id": 534207,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622341317.574, "dur": 3.809,
    "args": {
      "External id": 534208,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341320.151, "dur": 0.464,
    "args": {
      "External id": 534209,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622341385.471, "dur": 46.536,
    "args": {
      "External id": 534210,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622341436.390, "dur": 5.280,
    "args": {
      "External id": 534211,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341439.670, "dur": 1.147,
    "args": {
      "External id": 534212,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "25165824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341443.037, "dur": 22.460,
    "args": {
      "External id": 534213,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622341470.205, "dur": 6.375,
    "args": {
      "External id": 534214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622341472.195, "dur": 3.589,
    "args": {
      "External id": 534215,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341474.438, "dur": 1.177,
    "args": {
      "External id": 534216,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622341478.819, "dur": 39.989,
    "args": {
      "External id": 534217,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622341480.239, "dur": 38.012,
    "args": {
      "External id": 534218,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341522.502, "dur": 12.982,
    "args": {
      "External id": 534219,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622341540.544, "dur": 3.753,
    "args": {
      "External id": 534220,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341542.736, "dur": 0.743,
    "args": {
      "External id": 534221,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622341547.674, "dur": 47.685,
    "args": {
      "External id": 534222,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622341548.526, "dur": 4.229,
    "args": {
      "External id": 534223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622341549.419, "dur": 2.744,
    "args": {
      "External id": 534224,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341551.054, "dur": 0.914,
    "args": {
      "External id": 534225,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622341553.626, "dur": 41.366,
    "args": {
      "External id": 534226,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622341554.374, "dur": 40.114,
    "args": {
      "External id": 534227,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622341599.001, "dur": 5.676,
    "args": {
      "External id": 534228,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341600.719, "dur": 2.809,
    "args": {
      "External id": 534229,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622341609.877, "dur": 1.379,
    "args": {
      "External id": 534230,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622341618.083, "dur": 6.167,
    "args": {
      "External id": 534231,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622341620.403, "dur": 3.580,
    "args": {
      "External id": 534232,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622341763.812, "dur": 179.999,
    "args": {
      "External id": 534233,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622341766.466, "dur": 3.026,
    "args": {
      "External id": 534234,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622341771.954, "dur": 171.347,
    "args": {
      "External id": 534235,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622341773.492, "dur": 0.419,
    "args": {
      "External id": 534236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622341775.382, "dur": 24.732,
    "args": {
      "External id": 534237,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622341801.757, "dur": 3.841,
    "args": {
      "External id": 534238,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341804.012, "dur": 1.343,
    "args": {
      "External id": 534239,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622341806.573, "dur": 24.077,
    "args": {
      "External id": 534240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622341807.563, "dur": 1.768,
    "args": {
      "External id": 534241,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622341810.458, "dur": 19.874,
    "args": {
      "External id": 534242,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341815.719, "dur": 2.573,
    "args": {
      "External id": 534243,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622341832.123, "dur": 19.597,
    "args": {
      "External id": 534244,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341853.193, "dur": 13.229,
    "args": {
      "External id": 534245,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622341868.762, "dur": 13.252,
    "args": {
      "External id": 534246,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341883.507, "dur": 11.739,
    "args": {
      "External id": 534247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622341897.345, "dur": 19.582,
    "args": {
      "External id": 534248,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341899.044, "dur": 1.810,
    "args": {
      "External id": 534249,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341903.309, "dur": 0.876,
    "args": {
      "External id": 534250,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341918.670, "dur": 11.337,
    "args": {
      "External id": 534251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622341931.111, "dur": 10.820,
    "args": {
      "External id": 534252,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622341950.595, "dur": 2.397,
    "args": {
      "External id": 534253,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622341963.043, "dur": 3.632,
    "args": {
      "External id": 534254,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622341965.458, "dur": 0.416,
    "args": {
      "External id": 534255,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622342031.058, "dur": 49.726,
    "args": {
      "External id": 534256,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622342085.537, "dur": 5.696,
    "args": {
      "External id": 534257,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342088.796, "dur": 1.340,
    "args": {
      "External id": 534258,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342092.673, "dur": 41.256,
    "args": {
      "External id": 534259,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622342140.834, "dur": 6.636,
    "args": {
      "External id": 534260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622342142.887, "dur": 3.697,
    "args": {
      "External id": 534261,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342145.317, "dur": 1.013,
    "args": {
      "External id": 534262,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622342150.513, "dur": 49.022,
    "args": {
      "External id": 534263,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622342151.925, "dur": 46.894,
    "args": {
      "External id": 534264,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342204.190, "dur": 15.869,
    "args": {
      "External id": 534265,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622342226.256, "dur": 4.785,
    "args": {
      "External id": 534266,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342229.015, "dur": 0.962,
    "args": {
      "External id": 534267,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622342235.178, "dur": 52.342,
    "args": {
      "External id": 534268,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622342236.110, "dur": 6.954,
    "args": {
      "External id": 534269,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622342237.230, "dur": 5.033,
    "args": {
      "External id": 534270,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342241.436, "dur": 0.671,
    "args": {
      "External id": 534271,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622342243.785, "dur": 43.312,
    "args": {
      "External id": 534272,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622342244.823, "dur": 41.712,
    "args": {
      "External id": 534273,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622342291.577, "dur": 3.899,
    "args": {
      "External id": 534274,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342293.401, "dur": 0.874,
    "args": {
      "External id": 534275,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622342301.439, "dur": 1.574,
    "args": {
      "External id": 534276,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622342311.101, "dur": 6.711,
    "args": {
      "External id": 534277,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622342313.415, "dur": 4.126,
    "args": {
      "External id": 534278,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622342404.045, "dur": 172.658,
    "args": {
      "External id": 534279,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622342406.270, "dur": 2.113,
    "args": {
      "External id": 534280,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622342409.992, "dur": 166.306,
    "args": {
      "External id": 534281,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622342413.483, "dur": 0.187,
    "args": {
      "External id": 534282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622342415.240, "dur": 19.886,
    "args": {
      "External id": 534283,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622342436.639, "dur": 4.256,
    "args": {
      "External id": 534284,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342439.248, "dur": 1.331,
    "args": {
      "External id": 534285,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622342441.905, "dur": 23.476,
    "args": {
      "External id": 534286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622342445.365, "dur": 1.880,
    "args": {
      "External id": 534287,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622342448.335, "dur": 16.690,
    "args": {
      "External id": 534288,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342451.036, "dur": 2.621,
    "args": {
      "External id": 534289,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622342466.921, "dur": 22.139,
    "args": {
      "External id": 534290,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342490.458, "dur": 11.902,
    "args": {
      "External id": 534291,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622342504.809, "dur": 13.198,
    "args": {
      "External id": 534292,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342519.205, "dur": 11.200,
    "args": {
      "External id": 534293,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622342532.059, "dur": 18.684,
    "args": {
      "External id": 534294,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342533.927, "dur": 1.829,
    "args": {
      "External id": 534295,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342538.178, "dur": 0.925,
    "args": {
      "External id": 534296,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342552.365, "dur": 11.194,
    "args": {
      "External id": 534297,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342564.580, "dur": 10.369,
    "args": {
      "External id": 534298,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622342582.647, "dur": 1.681,
    "args": {
      "External id": 534299,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622342607.243, "dur": 3.877,
    "args": {
      "External id": 534300,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342609.618, "dur": 0.636,
    "args": {
      "External id": 534301,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622342711.971, "dur": 50.940,
    "args": {
      "External id": 534302,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622342768.400, "dur": 8.375,
    "args": {
      "External id": 534303,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342771.841, "dur": 3.407,
    "args": {
      "External id": 534304,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "41943040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342778.262, "dur": 33.891,
    "args": {
      "External id": 534305,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622342817.821, "dur": 6.290,
    "args": {
      "External id": 534306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622342819.437, "dur": 3.908,
    "args": {
      "External id": 534307,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342821.920, "dur": 1.082,
    "args": {
      "External id": 534308,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622342826.821, "dur": 68.362,
    "args": {
      "External id": 534309,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622342827.857, "dur": 66.719,
    "args": {
      "External id": 534310,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622342899.302, "dur": 38.159,
    "args": {
      "External id": 534311,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622342943.009, "dur": 4.100,
    "args": {
      "External id": 534312,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342945.383, "dur": 0.785,
    "args": {
      "External id": 534313,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622342951.074, "dur": 52.334,
    "args": {
      "External id": 534314,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622342952.027, "dur": 4.567,
    "args": {
      "External id": 534315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622342952.873, "dur": 3.164,
    "args": {
      "External id": 534316,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622342954.698, "dur": 1.114,
    "args": {
      "External id": 534317,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622342957.325, "dur": 45.655,
    "args": {
      "External id": 534318,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622342957.815, "dur": 44.544,
    "args": {
      "External id": 534319,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622343007.335, "dur": 3.967,
    "args": {
      "External id": 534320,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343009.250, "dur": 0.880,
    "args": {
      "External id": 534321,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622343017.371, "dur": 1.914,
    "args": {
      "External id": 534322,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343027.172, "dur": 6.982,
    "args": {
      "External id": 534323,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343029.286, "dur": 4.494,
    "args": {
      "External id": 534324,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622343136.353, "dur": 176.456,
    "args": {
      "External id": 534325,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343138.690, "dur": 5.104,
    "args": {
      "External id": 534326,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622343145.534, "dur": 166.599,
    "args": {
      "External id": 534327,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622343147.037, "dur": 0.522,
    "args": {
      "External id": 534328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622343148.991, "dur": 21.876,
    "args": {
      "External id": 534329,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622343172.609, "dur": 5.627,
    "args": {
      "External id": 534330,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343177.022, "dur": 0.936,
    "args": {
      "External id": 534331,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622343179.255, "dur": 20.363,
    "args": {
      "External id": 534332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343180.496, "dur": 1.382,
    "args": {
      "External id": 534333,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622343183.012, "dur": 16.358,
    "args": {
      "External id": 534334,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343185.824, "dur": 2.262,
    "args": {
      "External id": 534335,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622343201.357, "dur": 19.339,
    "args": {
      "External id": 534336,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343222.104, "dur": 13.257,
    "args": {
      "External id": 534337,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622343237.667, "dur": 13.656,
    "args": {
      "External id": 534338,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343252.563, "dur": 11.764,
    "args": {
      "External id": 534339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622343266.052, "dur": 20.713,
    "args": {
      "External id": 534340,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343270.273, "dur": 1.769,
    "args": {
      "External id": 534341,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343274.505, "dur": 0.972,
    "args": {
      "External id": 534342,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343288.066, "dur": 11.276,
    "args": {
      "External id": 534343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343300.634, "dur": 10.375,
    "args": {
      "External id": 534344,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622343319.265, "dur": 2.367,
    "args": {
      "External id": 534345,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622343330.346, "dur": 3.512,
    "args": {
      "External id": 534346,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343332.632, "dur": 0.362,
    "args": {
      "External id": 534347,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622343394.879, "dur": 43.546,
    "args": {
      "External id": 534348,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622343443.239, "dur": 4.872,
    "args": {
      "External id": 534349,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343446.220, "dur": 0.793,
    "args": {
      "External id": 534350,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343449.380, "dur": 21.998,
    "args": {
      "External id": 534351,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622343476.262, "dur": 5.854,
    "args": {
      "External id": 534352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622343477.978, "dur": 3.450,
    "args": {
      "External id": 534353,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343479.935, "dur": 1.314,
    "args": {
      "External id": 534354,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622343484.359, "dur": 40.439,
    "args": {
      "External id": 534355,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622343485.230, "dur": 38.920,
    "args": {
      "External id": 534356,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343528.172, "dur": 13.317,
    "args": {
      "External id": 534357,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622343547.090, "dur": 4.623,
    "args": {
      "External id": 534358,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343549.998, "dur": 0.799,
    "args": {
      "External id": 534359,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 4183438, "tid": 4183438,
    "ts": 679622343555.395, "dur": 45.015,
    "args": {
      "External id": 534360,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [32000, 1024], []], "Ev Idx": 8519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622343556.315, "dur": 3.795,
    "args": {
      "External id": 534361,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622343557.182, "dur": 2.387,
    "args": {
      "External id": 534362,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], []], "Input Dims": [[32000, 1024], [], []], "Ev Idx": 8521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343558.724, "dur": 0.663,
    "args": {
      "External id": 534363,"Record function id": 0, "Concrete Inputs": ["", "[1024, 32000]", "[1, 1024]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[32000, 1024], [], [], []], "Ev Idx": 8522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622343560.733, "dur": 39.314,
    "args": {
      "External id": 534364,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622343561.269, "dur": 38.239,
    "args": {
      "External id": 534365,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1024, 1], [1, 1024]], "Input Dims": [[8192, 1024], [1024, 32000]], "Ev Idx": 8524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622343603.957, "dur": 3.473,
    "args": {
      "External id": 534366,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343605.707, "dur": 0.670,
    "args": {
      "External id": 534367,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622343612.121, "dur": 1.623,
    "args": {
      "External id": 534368,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343620.980, "dur": 6.150,
    "args": {
      "External id": 534369,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343623.142, "dur": 3.659,
    "args": {
      "External id": 534370,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622343765.405, "dur": 177.118,
    "args": {
      "External id": 534371,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343769.992, "dur": 2.905,
    "args": {
      "External id": 534372,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 4183438, "tid": 4183438,
    "ts": 679622343774.388, "dur": 167.705,
    "args": {
      "External id": 534373,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 4183438, "tid": 4183438,
    "ts": 679622343776.090, "dur": 0.467,
    "args": {
      "External id": 534374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 4183438, "tid": 4183438,
    "ts": 679622343777.728, "dur": 20.935,
    "args": {
      "External id": 534375,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622343800.311, "dur": 6.014,
    "args": {
      "External id": 534376,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343804.751, "dur": 1.157,
    "args": {
      "External id": 534377,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622343807.191, "dur": 20.510,
    "args": {
      "External id": 534378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622343808.202, "dur": 1.403,
    "args": {
      "External id": 534379,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622343810.690, "dur": 16.719,
    "args": {
      "External id": 534380,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343813.276, "dur": 2.829,
    "args": {
      "External id": 534381,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622343829.273, "dur": 21.258,
    "args": {
      "External id": 534382,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343851.981, "dur": 14.465,
    "args": {
      "External id": 534383,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 4183438, "tid": 4183438,
    "ts": 679622343869.233, "dur": 14.043,
    "args": {
      "External id": 534384,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343884.785, "dur": 11.810,
    "args": {
      "External id": 534385,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622343898.354, "dur": 17.985,
    "args": {
      "External id": 534386,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343900.272, "dur": 1.635,
    "args": {
      "External id": 534387,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343904.080, "dur": 0.927,
    "args": {
      "External id": 534388,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343917.917, "dur": 10.952,
    "args": {
      "External id": 534389,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622343929.978, "dur": 10.865,
    "args": {
      "External id": 534390,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622343949.190, "dur": 2.033,
    "args": {
      "External id": 534391,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622343960.046, "dur": 3.865,
    "args": {
      "External id": 534392,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622343962.336, "dur": 0.641,
    "args": {
      "External id": 534393,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622344022.833, "dur": 46.419,
    "args": {
      "External id": 534394,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [1024, 1]], "Input Dims": [[8192, 32000], [32000, 1024]], "Ev Idx": 8553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 4183438, "tid": 4183438,
    "ts": 679622344073.860, "dur": 14.022,
    "args": {
      "External id": 534395,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1024, 1], [], [], [], []], "Input Dims": [[65536, 1024], [], [], [], []], "Ev Idx": 8554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622344085.836, "dur": 0.918,
    "args": {
      "External id": 534396,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1024]", "[1024, 1]", "58720256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1024, 1], [], [], []], "Input Dims": [[65536, 1024], [], [], []], "Ev Idx": 8555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622344089.456, "dur": 40.424,
    "args": {
      "External id": 534397,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8192, 1024], [8192, 1024], []], "Ev Idx": 8556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 4183438, "tid": 4183438,
    "ts": 679622344136.348, "dur": 9.016,
    "args": {
      "External id": 534398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 4183438, "tid": 4183438,
    "ts": 679622344138.120, "dur": 6.440,
    "args": {
      "External id": 534399,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622344140.924, "dur": 3.416,
    "args": {
      "External id": 534400,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 4183438, "tid": 4183438,
    "ts": 679622344148.522, "dur": 45.691,
    "args": {
      "External id": 534401,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 4183438, "tid": 4183438,
    "ts": 679622344149.787, "dur": 43.789,
    "args": {
      "External id": 534402,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [1024, 1]], "Input Dims": [[32000, 8192], [8192, 1024]], "Ev Idx": 8561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622344197.966, "dur": 16.998,
    "args": {
      "External id": 534403,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622344219.581, "dur": 25.974,
    "args": {
      "External id": 534404,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 4183438, "tid": 4183438,
    "ts": 679622344222.258, "dur": 22.913,
    "args": {
      "External id": 534405,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622344227.257, "dur": 1.385,
    "args": {
      "External id": 534406,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622344250.725, "dur": 27.853,
    "args": {
      "External id": 534407,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], []], "Ev Idx": 8566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622344252.630, "dur": 25.751,
    "args": {
      "External id": 534408,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1024, 1], [], [], [], [], [], []], "Input Dims": [[32000, 1024], [], [], [], [], [], []], "Ev Idx": 8567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622344257.495, "dur": 4.017,
    "args": {
      "External id": 534409,"Record function id": 0, "Concrete Inputs": ["[32000, 1024]", "[1024, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622344262.671, "dur": 15.141,
    "args": {
      "External id": 534410,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[32000, 1024], [32000, 1024], []], "Ev Idx": 8569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622344291.205, "dur": 5.387,
    "args": {
      "External id": 534411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 8570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622344293.188, "dur": 3.069,
    "args": {
      "External id": 534412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[65536, 1024]], "Ev Idx": 8571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622344297.564, "dur": 1.518,
    "args": {
      "External id": 534413,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622344298.392, "dur": 0.623,
    "args": {
      "External id": 534414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1024, 1]], "Input Dims": [[32000, 1024]], "Ev Idx": 8573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622344342.332, "dur": 22.031,
    "args": {
      "External id": 534415,"Sequence number": 6319628, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 8574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622344366.784, "dur": 15.035,
    "args": {
      "External id": 534416,"Sequence number": 6319629, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 8575
    }
  },
  {
    "ph": "s", "id": 2, "pid": 4183438, "tid": 4183438, "ts": 679622344366.784,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward", "pid": 4183438, "tid": 4183438,
    "ts": 679622344490.915, "dur": 39.363,
    "args": {
      "External id": 534417,"Record function id": 0, "Ev Idx": 8576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 4183438, "tid": 4183438,
    "ts": 679622344629.456, "dur": 81.363,
    "args": {
      "External id": 534418,"Sequence number": 6319630, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 8577
    }
  },
  {
    "ph": "s", "id": 1, "pid": 4183438, "tid": 4183438, "ts": 679622344629.456,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622344771.349, "dur": 35.624,
    "args": {
      "External id": 534419,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 8578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622344773.124, "dur": 9.710,
    "args": {
      "External id": 534420,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 8579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622344776.954, "dur": 5.221,
    "args": {
      "External id": 534421,"Record function id": 0, "Concrete Inputs": ["[1]", "[1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622344784.445, "dur": 22.178,
    "args": {
      "External id": 534422,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 8581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 4183438, "tid": 4183438,
    "ts": 679622808427.349, "dur": 77.936,
    "args": {
      "External id": 534423,"Sequence number": 6319631, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 4183438, "tid": 4183438,
    "ts": 679622808515.733, "dur": 21.881,
    "args": {
      "External id": 534424,"Sequence number": 6319632, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 8583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622808622.547, "dur": 21.682,
    "args": {
      "External id": 534425,"Sequence number": 6319633, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "long int", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 8584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622809081.532, "dur": 45.850,
    "args": {
      "External id": 534426,"Sequence number": 6319634, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 8585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622809136.905, "dur": 15.395,
    "args": {
      "External id": 534427,"Sequence number": 6319635, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "long int", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 8586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_norm", "pid": 4183438, "tid": 4183438,
    "ts": 679622810491.121, "dur": 2305.088,
    "args": {
      "External id": 534428,"Record function id": 0, "Concrete Inputs": ["", "2.", ""], "Input type": ["TensorList", "Scalar", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 8587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_norm", "pid": 4183438, "tid": 4183438,
    "ts": 679622810972.688, "dur": 626.954,
    "args": {
      "External id": 534429,"Record function id": 0, "Concrete Inputs": ["", "2.", ""], "Input type": ["TensorList", "Scalar", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 8588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 4183438, "tid": 4183438,
    "ts": 679622810991.037, "dur": 61.678,
    "args": {
      "External id": 534430,"Record function id": 0, "Concrete Inputs": ["[13797]", "6", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 8589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622810994.296, "dur": 11.979,
    "args": {
      "External id": 534431,"Record function id": 0, "Concrete Inputs": ["[13797]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 4183438, "tid": 4183438,
    "ts": 679622811009.040, "dur": 43.338,
    "args": {
      "External id": 534432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[13797]], "Ev Idx": 8591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622811011.284, "dur": 40.211,
    "args": {
      "External id": 534433,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[13797], []], "Ev Idx": 8592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812822.825, "dur": 3.141,
    "args": {
      "External id": 534434,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812827.995, "dur": 0.418,
    "args": {
      "External id": 534435,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812829.683, "dur": 0.382,
    "args": {
      "External id": 534436,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812831.111, "dur": 0.320,
    "args": {
      "External id": 534437,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812832.599, "dur": 0.382,
    "args": {
      "External id": 534438,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812834.089, "dur": 0.378,
    "args": {
      "External id": 534439,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812835.465, "dur": 0.381,
    "args": {
      "External id": 534440,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812838.059, "dur": 0.517,
    "args": {
      "External id": 534441,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812839.500, "dur": 0.376,
    "args": {
      "External id": 534442,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812840.837, "dur": 0.334,
    "args": {
      "External id": 534443,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812842.011, "dur": 0.221,
    "args": {
      "External id": 534444,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812843.069, "dur": 0.231,
    "args": {
      "External id": 534445,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812844.208, "dur": 0.215,
    "args": {
      "External id": 534446,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812845.286, "dur": 0.212,
    "args": {
      "External id": 534447,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812846.422, "dur": 0.217,
    "args": {
      "External id": 534448,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812848.844, "dur": 0.224,
    "args": {
      "External id": 534449,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812849.918, "dur": 0.226,
    "args": {
      "External id": 534450,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812854.510, "dur": 0.219,
    "args": {
      "External id": 534451,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812855.614, "dur": 0.197,
    "args": {
      "External id": 534452,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812856.673, "dur": 0.190,
    "args": {
      "External id": 534453,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812857.671, "dur": 0.246,
    "args": {
      "External id": 534454,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812858.775, "dur": 0.222,
    "args": {
      "External id": 534455,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812859.942, "dur": 0.226,
    "args": {
      "External id": 534456,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812862.351, "dur": 0.333,
    "args": {
      "External id": 534457,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812863.574, "dur": 0.318,
    "args": {
      "External id": 534458,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812864.882, "dur": 0.218,
    "args": {
      "External id": 534459,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812865.996, "dur": 0.233,
    "args": {
      "External id": 534460,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812867.044, "dur": 0.253,
    "args": {
      "External id": 534461,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812868.133, "dur": 0.235,
    "args": {
      "External id": 534462,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812869.379, "dur": 0.246,
    "args": {
      "External id": 534463,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812870.477, "dur": 0.236,
    "args": {
      "External id": 534464,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812873.059, "dur": 0.331,
    "args": {
      "External id": 534465,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812874.187, "dur": 0.352,
    "args": {
      "External id": 534466,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812875.685, "dur": 0.239,
    "args": {
      "External id": 534467,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812876.890, "dur": 0.209,
    "args": {
      "External id": 534468,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812878.166, "dur": 0.338,
    "args": {
      "External id": 534469,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812879.359, "dur": 0.352,
    "args": {
      "External id": 534470,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812880.569, "dur": 0.281,
    "args": {
      "External id": 534471,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812881.648, "dur": 0.326,
    "args": {
      "External id": 534472,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812883.805, "dur": 0.354,
    "args": {
      "External id": 534473,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812884.974, "dur": 0.387,
    "args": {
      "External id": 534474,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812886.339, "dur": 0.201,
    "args": {
      "External id": 534475,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812887.384, "dur": 0.333,
    "args": {
      "External id": 534476,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812888.510, "dur": 0.362,
    "args": {
      "External id": 534477,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812889.702, "dur": 0.295,
    "args": {
      "External id": 534478,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812891.021, "dur": 0.218,
    "args": {
      "External id": 534479,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812892.067, "dur": 0.202,
    "args": {
      "External id": 534480,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812895.946, "dur": 0.266,
    "args": {
      "External id": 534481,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812897.116, "dur": 0.347,
    "args": {
      "External id": 534482,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812898.330, "dur": 0.352,
    "args": {
      "External id": 534483,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812899.478, "dur": 0.339,
    "args": {
      "External id": 534484,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812900.659, "dur": 0.200,
    "args": {
      "External id": 534485,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812901.655, "dur": 0.234,
    "args": {
      "External id": 534486,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812902.943, "dur": 0.203,
    "args": {
      "External id": 534487,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812904.043, "dur": 0.202,
    "args": {
      "External id": 534488,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812906.289, "dur": 0.200,
    "args": {
      "External id": 534489,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812907.296, "dur": 0.198,
    "args": {
      "External id": 534490,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812908.318, "dur": 0.199,
    "args": {
      "External id": 534491,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812912.584, "dur": 0.194,
    "args": {
      "External id": 534492,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812913.577, "dur": 0.200,
    "args": {
      "External id": 534493,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812914.622, "dur": 0.200,
    "args": {
      "External id": 534494,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812915.619, "dur": 0.203,
    "args": {
      "External id": 534495,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812916.615, "dur": 0.199,
    "args": {
      "External id": 534496,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812919.044, "dur": 0.207,
    "args": {
      "External id": 534497,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812920.089, "dur": 0.221,
    "args": {
      "External id": 534498,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812921.403, "dur": 0.255,
    "args": {
      "External id": 534499,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812922.546, "dur": 0.253,
    "args": {
      "External id": 534500,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812923.653, "dur": 0.209,
    "args": {
      "External id": 534501,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812924.717, "dur": 0.211,
    "args": {
      "External id": 534502,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812925.997, "dur": 0.206,
    "args": {
      "External id": 534503,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812927.071, "dur": 0.318,
    "args": {
      "External id": 534504,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812929.466, "dur": 0.323,
    "args": {
      "External id": 534505,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812930.632, "dur": 0.238,
    "args": {
      "External id": 534506,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812931.815, "dur": 0.212,
    "args": {
      "External id": 534507,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812932.856, "dur": 0.241,
    "args": {
      "External id": 534508,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812933.977, "dur": 0.216,
    "args": {
      "External id": 534509,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812935.027, "dur": 0.215,
    "args": {
      "External id": 534510,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812936.706, "dur": 0.235,
    "args": {
      "External id": 534511,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812937.769, "dur": 0.210,
    "args": {
      "External id": 534512,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812940.039, "dur": 0.242,
    "args": {
      "External id": 534513,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812941.173, "dur": 0.204,
    "args": {
      "External id": 534514,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812942.198, "dur": 0.237,
    "args": {
      "External id": 534515,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812943.223, "dur": 0.198,
    "args": {
      "External id": 534516,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812944.218, "dur": 0.215,
    "args": {
      "External id": 534517,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812945.253, "dur": 0.251,
    "args": {
      "External id": 534518,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812946.338, "dur": 0.227,
    "args": {
      "External id": 534519,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812947.443, "dur": 0.195,
    "args": {
      "External id": 534520,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812949.449, "dur": 0.204,
    "args": {
      "External id": 534521,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812950.484, "dur": 0.203,
    "args": {
      "External id": 534522,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812951.489, "dur": 0.192,
    "args": {
      "External id": 534523,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812952.548, "dur": 0.198,
    "args": {
      "External id": 534524,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812953.546, "dur": 0.193,
    "args": {
      "External id": 534525,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812954.548, "dur": 0.200,
    "args": {
      "External id": 534526,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812956.170, "dur": 0.205,
    "args": {
      "External id": 534527,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812957.215, "dur": 0.198,
    "args": {
      "External id": 534528,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812978.177, "dur": 0.224,
    "args": {
      "External id": 539137,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812979.480, "dur": 0.205,
    "args": {
      "External id": 539138,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812980.528, "dur": 0.204,
    "args": {
      "External id": 539139,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812981.542, "dur": 0.195,
    "args": {
      "External id": 539140,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812982.721, "dur": 0.218,
    "args": {
      "External id": 539141,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812983.751, "dur": 0.202,
    "args": {
      "External id": 539142,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812984.961, "dur": 0.207,
    "args": {
      "External id": 539143,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812986.032, "dur": 0.199,
    "args": {
      "External id": 539144,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812988.604, "dur": 0.236,
    "args": {
      "External id": 539145,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812989.670, "dur": 0.187,
    "args": {
      "External id": 539146,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812990.659, "dur": 0.192,
    "args": {
      "External id": 539147,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812991.720, "dur": 0.187,
    "args": {
      "External id": 539148,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812992.813, "dur": 0.197,
    "args": {
      "External id": 539149,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812993.859, "dur": 0.211,
    "args": {
      "External id": 539150,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812998.624, "dur": 0.210,
    "args": {
      "External id": 539151,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622812999.647, "dur": 0.236,
    "args": {
      "External id": 539152,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813002.179, "dur": 0.202,
    "args": {
      "External id": 539153,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813003.179, "dur": 0.197,
    "args": {
      "External id": 539154,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813004.199, "dur": 0.201,
    "args": {
      "External id": 539155,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813005.277, "dur": 0.193,
    "args": {
      "External id": 539156,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813006.281, "dur": 0.198,
    "args": {
      "External id": 539157,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813007.387, "dur": 0.187,
    "args": {
      "External id": 539158,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813008.457, "dur": 0.196,
    "args": {
      "External id": 539159,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813013.032, "dur": 0.205,
    "args": {
      "External id": 539160,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813015.557, "dur": 0.207,
    "args": {
      "External id": 539161,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813016.524, "dur": 0.197,
    "args": {
      "External id": 539162,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813017.503, "dur": 0.216,
    "args": {
      "External id": 539163,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813018.533, "dur": 0.203,
    "args": {
      "External id": 539164,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813019.568, "dur": 0.217,
    "args": {
      "External id": 539165,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813020.653, "dur": 0.201,
    "args": {
      "External id": 539166,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813021.685, "dur": 0.202,
    "args": {
      "External id": 539167,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813022.777, "dur": 0.189,
    "args": {
      "External id": 539168,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813024.942, "dur": 0.207,
    "args": {
      "External id": 539169,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813025.973, "dur": 0.197,
    "args": {
      "External id": 539170,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813027.437, "dur": 0.210,
    "args": {
      "External id": 539171,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813028.513, "dur": 0.197,
    "args": {
      "External id": 539172,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813029.533, "dur": 0.200,
    "args": {
      "External id": 539173,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813030.572, "dur": 0.188,
    "args": {
      "External id": 539174,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813031.566, "dur": 0.195,
    "args": {
      "External id": 539175,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813032.567, "dur": 0.187,
    "args": {
      "External id": 539176,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813034.610, "dur": 0.210,
    "args": {
      "External id": 539177,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813035.668, "dur": 0.198,
    "args": {
      "External id": 539178,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813036.663, "dur": 0.199,
    "args": {
      "External id": 539179,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813037.673, "dur": 0.193,
    "args": {
      "External id": 539180,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813038.788, "dur": 0.198,
    "args": {
      "External id": 539181,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813039.781, "dur": 0.189,
    "args": {
      "External id": 539182,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813040.779, "dur": 0.200,
    "args": {
      "External id": 539183,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813041.806, "dur": 0.197,
    "args": {
      "External id": 539184,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813044.163, "dur": 0.203,
    "args": {
      "External id": 539185,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813045.201, "dur": 0.200,
    "args": {
      "External id": 539186,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813046.242, "dur": 0.194,
    "args": {
      "External id": 539187,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813047.402, "dur": 0.191,
    "args": {
      "External id": 539188,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813048.490, "dur": 0.198,
    "args": {
      "External id": 539189,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813049.533, "dur": 0.197,
    "args": {
      "External id": 539190,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813051.121, "dur": 0.207,
    "args": {
      "External id": 539191,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813052.159, "dur": 0.193,
    "args": {
      "External id": 539192,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813054.326, "dur": 0.202,
    "args": {
      "External id": 539193,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813055.333, "dur": 0.197,
    "args": {
      "External id": 539194,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813056.334, "dur": 0.196,
    "args": {
      "External id": 539195,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813057.408, "dur": 0.201,
    "args": {
      "External id": 539196,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813058.403, "dur": 0.199,
    "args": {
      "External id": 539197,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813059.416, "dur": 0.187,
    "args": {
      "External id": 539198,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813060.447, "dur": 0.196,
    "args": {
      "External id": 539199,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813061.508, "dur": 0.203,
    "args": {
      "External id": 539200,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813064.037, "dur": 0.202,
    "args": {
      "External id": 539201,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813065.026, "dur": 0.204,
    "args": {
      "External id": 539202,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813067.550, "dur": 0.216,
    "args": {
      "External id": 539203,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813068.659, "dur": 0.204,
    "args": {
      "External id": 539204,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813069.704, "dur": 0.196,
    "args": {
      "External id": 539205,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813070.700, "dur": 0.196,
    "args": {
      "External id": 539206,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813071.701, "dur": 0.198,
    "args": {
      "External id": 539207,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813072.733, "dur": 0.199,
    "args": {
      "External id": 539208,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813075.411, "dur": 0.214,
    "args": {
      "External id": 539209,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813076.429, "dur": 0.200,
    "args": {
      "External id": 539210,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813077.461, "dur": 0.203,
    "args": {
      "External id": 539211,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813078.580, "dur": 0.192,
    "args": {
      "External id": 539212,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813079.658, "dur": 0.207,
    "args": {
      "External id": 539213,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813080.700, "dur": 0.204,
    "args": {
      "External id": 539214,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813082.081, "dur": 0.206,
    "args": {
      "External id": 539215,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813083.134, "dur": 0.198,
    "args": {
      "External id": 539216,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813085.385, "dur": 0.245,
    "args": {
      "External id": 539217,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813086.449, "dur": 0.204,
    "args": {
      "External id": 539218,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813087.652, "dur": 0.196,
    "args": {
      "External id": 539219,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813088.697, "dur": 0.201,
    "args": {
      "External id": 539220,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813089.718, "dur": 0.203,
    "args": {
      "External id": 539221,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813090.744, "dur": 0.196,
    "args": {
      "External id": 539222,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813091.779, "dur": 0.193,
    "args": {
      "External id": 539223,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813092.784, "dur": 0.188,
    "args": {
      "External id": 539224,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813095.038, "dur": 0.207,
    "args": {
      "External id": 539225,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813096.098, "dur": 0.207,
    "args": {
      "External id": 539226,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813097.108, "dur": 0.199,
    "args": {
      "External id": 539227,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813098.125, "dur": 0.202,
    "args": {
      "External id": 539228,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813099.165, "dur": 0.245,
    "args": {
      "External id": 539229,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813113.159, "dur": 0.608,
    "args": {
      "External id": 539230,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813114.794, "dur": 0.208,
    "args": {
      "External id": 539231,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813115.880, "dur": 0.227,
    "args": {
      "External id": 539232,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813118.488, "dur": 0.213,
    "args": {
      "External id": 539233,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813119.523, "dur": 0.233,
    "args": {
      "External id": 539234,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813120.621, "dur": 0.218,
    "args": {
      "External id": 539235,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813121.824, "dur": 0.207,
    "args": {
      "External id": 539236,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813122.928, "dur": 0.204,
    "args": {
      "External id": 539237,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813123.949, "dur": 0.230,
    "args": {
      "External id": 539238,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813125.081, "dur": 0.200,
    "args": {
      "External id": 539239,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813126.069, "dur": 0.211,
    "args": {
      "External id": 539240,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813128.204, "dur": 0.200,
    "args": {
      "External id": 539241,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813129.208, "dur": 0.204,
    "args": {
      "External id": 539242,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813130.597, "dur": 0.215,
    "args": {
      "External id": 539243,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813131.772, "dur": 0.242,
    "args": {
      "External id": 539244,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813138.231, "dur": 0.227,
    "args": {
      "External id": 539245,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813139.415, "dur": 0.207,
    "args": {
      "External id": 539246,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813140.389, "dur": 0.212,
    "args": {
      "External id": 539247,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813141.404, "dur": 0.202,
    "args": {
      "External id": 539248,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813143.910, "dur": 0.222,
    "args": {
      "External id": 539249,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813144.935, "dur": 0.206,
    "args": {
      "External id": 539250,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813145.928, "dur": 0.197,
    "args": {
      "External id": 539251,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813146.936, "dur": 0.193,
    "args": {
      "External id": 539252,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813147.951, "dur": 0.200,
    "args": {
      "External id": 539253,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813148.934, "dur": 0.230,
    "args": {
      "External id": 539254,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813149.950, "dur": 0.207,
    "args": {
      "External id": 539255,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813150.938, "dur": 0.206,
    "args": {
      "External id": 539256,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813153.221, "dur": 0.220,
    "args": {
      "External id": 539257,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813154.230, "dur": 0.208,
    "args": {
      "External id": 539258,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813155.253, "dur": 0.204,
    "args": {
      "External id": 539259,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622813156.244, "dur": 0.193,
    "args": {
      "External id": 539260,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 8811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 4183438, "tid": 4183438,
    "ts": 679622813201.747, "dur": 1090.807,
    "args": {
      "External id": 539261,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 4183438, "tid": 4183438,
    "ts": 679622813544.207, "dur": 682.413,
    "args": {
      "External id": 539262,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813550.083, "dur": 7.111,
    "args": {
      "External id": 539263,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813554.076, "dur": 2.759,
    "args": {
      "External id": 539264,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813557.756, "dur": 2.888,
    "args": {
      "External id": 539265,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813558.409, "dur": 2.103,
    "args": {
      "External id": 539266,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813561.038, "dur": 2.435,
    "args": {
      "External id": 539267,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813562.942, "dur": 0.461,
    "args": {
      "External id": 539268,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813563.851, "dur": 0.896,
    "args": {
      "External id": 539269,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813564.248, "dur": 0.432,
    "args": {
      "External id": 539270,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813565.991, "dur": 4.289,
    "args": {
      "External id": 539271,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813569.667, "dur": 0.550,
    "args": {
      "External id": 539272,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813570.623, "dur": 1.219,
    "args": {
      "External id": 539273,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813571.048, "dur": 0.730,
    "args": {
      "External id": 539274,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813572.081, "dur": 1.230,
    "args": {
      "External id": 539275,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813572.453, "dur": 0.794,
    "args": {
      "External id": 539276,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813573.543, "dur": 4.596,
    "args": {
      "External id": 539277,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813577.388, "dur": 0.682,
    "args": {
      "External id": 539278,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813578.711, "dur": 1.301,
    "args": {
      "External id": 539279,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813579.155, "dur": 0.793,
    "args": {
      "External id": 539280,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813580.231, "dur": 2.727,
    "args": {
      "External id": 539281,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813580.584, "dur": 2.284,
    "args": {
      "External id": 539282,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813583.218, "dur": 2.419,
    "args": {
      "External id": 539283,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813584.935, "dur": 0.633,
    "args": {
      "External id": 539284,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813585.949, "dur": 1.177,
    "args": {
      "External id": 539285,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813586.302, "dur": 0.759,
    "args": {
      "External id": 539286,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813588.227, "dur": 4.492,
    "args": {
      "External id": 539287,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813592.088, "dur": 0.565,
    "args": {
      "External id": 539288,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813592.984, "dur": 1.223,
    "args": {
      "External id": 539289,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813593.353, "dur": 0.789,
    "args": {
      "External id": 539290,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813594.674, "dur": 1.152,
    "args": {
      "External id": 539291,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813595.060, "dur": 0.699,
    "args": {
      "External id": 539292,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813596.045, "dur": 3.428,
    "args": {
      "External id": 539293,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813598.760, "dur": 0.647,
    "args": {
      "External id": 539294,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813599.887, "dur": 1.072,
    "args": {
      "External id": 539295,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813600.279, "dur": 0.611,
    "args": {
      "External id": 539296,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813601.180, "dur": 2.658,
    "args": {
      "External id": 539297,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813601.536, "dur": 2.230,
    "args": {
      "External id": 539298,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813604.059, "dur": 2.387,
    "args": {
      "External id": 539299,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813605.888, "dur": 0.492,
    "args": {
      "External id": 539300,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813606.765, "dur": 0.927,
    "args": {
      "External id": 539301,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813607.124, "dur": 0.501,
    "args": {
      "External id": 539302,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813608.831, "dur": 4.519,
    "args": {
      "External id": 539303,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813612.651, "dur": 0.628,
    "args": {
      "External id": 539304,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813613.592, "dur": 1.145,
    "args": {
      "External id": 539305,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813613.987, "dur": 0.681,
    "args": {
      "External id": 539306,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813614.958, "dur": 0.960,
    "args": {
      "External id": 539307,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813615.325, "dur": 0.526,
    "args": {
      "External id": 539308,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813616.135, "dur": 3.481,
    "args": {
      "External id": 539309,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813618.774, "dur": 0.775,
    "args": {
      "External id": 539310,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813619.835, "dur": 0.862,
    "args": {
      "External id": 539311,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813620.200, "dur": 0.429,
    "args": {
      "External id": 539312,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813620.965, "dur": 2.587,
    "args": {
      "External id": 539313,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813621.364, "dur": 2.113,
    "args": {
      "External id": 539314,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813623.777, "dur": 2.691,
    "args": {
      "External id": 539315,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813625.672, "dur": 0.715,
    "args": {
      "External id": 539316,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813626.775, "dur": 0.910,
    "args": {
      "External id": 539317,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813627.126, "dur": 0.491,
    "args": {
      "External id": 539318,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813628.904, "dur": 4.069,
    "args": {
      "External id": 539319,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813632.502, "dur": 0.403,
    "args": {
      "External id": 539320,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813633.213, "dur": 0.939,
    "args": {
      "External id": 539321,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813633.564, "dur": 0.521,
    "args": {
      "External id": 539322,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813634.388, "dur": 1.149,
    "args": {
      "External id": 539323,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813634.751, "dur": 0.721,
    "args": {
      "External id": 539324,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813635.778, "dur": 2.868,
    "args": {
      "External id": 539325,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813637.988, "dur": 0.587,
    "args": {
      "External id": 539326,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813638.859, "dur": 0.988,
    "args": {
      "External id": 539327,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813639.239, "dur": 0.540,
    "args": {
      "External id": 539328,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813640.071, "dur": 3.107,
    "args": {
      "External id": 539329,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813640.448, "dur": 2.658,
    "args": {
      "External id": 539330,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813643.405, "dur": 2.320,
    "args": {
      "External id": 539331,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813645.335, "dur": 0.327,
    "args": {
      "External id": 539332,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813645.944, "dur": 1.145,
    "args": {
      "External id": 539333,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813646.321, "dur": 0.704,
    "args": {
      "External id": 539334,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813647.587, "dur": 4.579,
    "args": {
      "External id": 539335,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813651.506, "dur": 0.596,
    "args": {
      "External id": 539336,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813652.455, "dur": 26.316,
    "args": {
      "External id": 539337,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813652.803, "dur": 0.245,
    "args": {
      "External id": 539338,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813680.598, "dur": 1.399,
    "args": {
      "External id": 539339,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813681.447, "dur": 0.461,
    "args": {
      "External id": 539340,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813682.423, "dur": 3.001,
    "args": {
      "External id": 539341,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813685.076, "dur": 0.273,
    "args": {
      "External id": 539342,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813685.697, "dur": 1.500,
    "args": {
      "External id": 539343,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813686.184, "dur": 0.948,
    "args": {
      "External id": 539344,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813687.544, "dur": 2.820,
    "args": {
      "External id": 539345,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813687.947, "dur": 2.342,
    "args": {
      "External id": 539346,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813690.912, "dur": 0.960,
    "args": {
      "External id": 539347,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813691.557, "dur": 0.250,
    "args": {
      "External id": 539348,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813692.128, "dur": 0.961,
    "args": {
      "External id": 539349,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813692.526, "dur": 0.498,
    "args": {
      "External id": 539350,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813693.384, "dur": 3.328,
    "args": {
      "External id": 539351,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813696.190, "dur": 0.456,
    "args": {
      "External id": 539352,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813697.064, "dur": 1.010,
    "args": {
      "External id": 539353,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813697.450, "dur": 0.556,
    "args": {
      "External id": 539354,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813698.332, "dur": 0.782,
    "args": {
      "External id": 539355,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813698.708, "dur": 0.339,
    "args": {
      "External id": 539356,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813699.372, "dur": 2.595,
    "args": {
      "External id": 539357,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813701.596, "dur": 0.305,
    "args": {
      "External id": 539358,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813702.206, "dur": 0.950,
    "args": {
      "External id": 539359,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813702.588, "dur": 0.503,
    "args": {
      "External id": 539360,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813703.489, "dur": 3.263,
    "args": {
      "External id": 539361,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813703.839, "dur": 2.652,
    "args": {
      "External id": 539362,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813707.023, "dur": 1.079,
    "args": {
      "External id": 539363,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813707.540, "dur": 0.490,
    "args": {
      "External id": 539364,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813708.346, "dur": 0.949,
    "args": {
      "External id": 539365,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813708.703, "dur": 0.526,
    "args": {
      "External id": 539366,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813709.735, "dur": 3.677,
    "args": {
      "External id": 539367,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813712.376, "dur": 0.967,
    "args": {
      "External id": 539368,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813713.688, "dur": 1.004,
    "args": {
      "External id": 539369,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813714.056, "dur": 0.567,
    "args": {
      "External id": 539370,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813715.137, "dur": 1.276,
    "args": {
      "External id": 539371,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813715.500, "dur": 0.847,
    "args": {
      "External id": 539372,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813716.857, "dur": 3.086,
    "args": {
      "External id": 539373,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813719.520, "dur": 0.357,
    "args": {
      "External id": 539374,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813720.202, "dur": 0.830,
    "args": {
      "External id": 539375,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813720.564, "dur": 0.399,
    "args": {
      "External id": 539376,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813721.294, "dur": 2.382,
    "args": {
      "External id": 539377,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813721.655, "dur": 1.945,
    "args": {
      "External id": 539378,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813723.929, "dur": 1.425,
    "args": {
      "External id": 539379,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813724.503, "dur": 0.787,
    "args": {
      "External id": 539380,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813725.694, "dur": 1.193,
    "args": {
      "External id": 539381,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813726.096, "dur": 0.719,
    "args": {
      "External id": 539382,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813727.162, "dur": 3.300,
    "args": {
      "External id": 539383,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813729.590, "dur": 0.805,
    "args": {
      "External id": 539384,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813730.746, "dur": 1.007,
    "args": {
      "External id": 539385,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813731.106, "dur": 0.577,
    "args": {
      "External id": 539386,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813732.156, "dur": 1.623,
    "args": {
      "External id": 539387,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813732.509, "dur": 1.204,
    "args": {
      "External id": 539388,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813734.147, "dur": 3.392,
    "args": {
      "External id": 539389,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813736.705, "dur": 0.769,
    "args": {
      "External id": 539390,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813737.934, "dur": 1.115,
    "args": {
      "External id": 539391,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813738.286, "dur": 0.697,
    "args": {
      "External id": 539392,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813739.294, "dur": 2.844,
    "args": {
      "External id": 539393,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813739.679, "dur": 2.390,
    "args": {
      "External id": 539394,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813742.381, "dur": 1.480,
    "args": {
      "External id": 539395,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813742.901, "dur": 0.891,
    "args": {
      "External id": 539396,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813744.250, "dur": 1.432,
    "args": {
      "External id": 539397,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813744.918, "dur": 0.698,
    "args": {
      "External id": 539398,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813745.962, "dur": 3.608,
    "args": {
      "External id": 539399,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813748.794, "dur": 0.701,
    "args": {
      "External id": 539400,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813749.846, "dur": 1.024,
    "args": {
      "External id": 539401,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813750.293, "dur": 0.508,
    "args": {
      "External id": 539402,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813751.326, "dur": 1.290,
    "args": {
      "External id": 539403,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813751.962, "dur": 0.588,
    "args": {
      "External id": 539404,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813752.883, "dur": 3.484,
    "args": {
      "External id": 539405,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813755.722, "dur": 0.577,
    "args": {
      "External id": 539406,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813756.715, "dur": 1.297,
    "args": {
      "External id": 539407,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813757.542, "dur": 0.404,
    "args": {
      "External id": 539408,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813758.483, "dur": 2.922,
    "args": {
      "External id": 539409,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813758.924, "dur": 2.409,
    "args": {
      "External id": 539410,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813761.676, "dur": 1.011,
    "args": {
      "External id": 539411,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813762.308, "dur": 0.311,
    "args": {
      "External id": 539412,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813762.948, "dur": 0.954,
    "args": {
      "External id": 539413,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813763.368, "dur": 0.467,
    "args": {
      "External id": 539414,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813764.178, "dur": 3.089,
    "args": {
      "External id": 539415,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813766.438, "dur": 0.761,
    "args": {
      "External id": 539416,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813767.551, "dur": 1.740,
    "args": {
      "External id": 539417,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813768.535, "dur": 0.689,
    "args": {
      "External id": 539418,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813769.788, "dur": 1.322,
    "args": {
      "External id": 539419,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813770.599, "dur": 0.443,
    "args": {
      "External id": 539420,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813771.466, "dur": 2.940,
    "args": {
      "External id": 539421,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813774.108, "dur": 0.230,
    "args": {
      "External id": 539422,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813774.770, "dur": 0.834,
    "args": {
      "External id": 539423,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813775.171, "dur": 0.364,
    "args": {
      "External id": 539424,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813775.866, "dur": 3.227,
    "args": {
      "External id": 539425,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813776.350, "dur": 2.669,
    "args": {
      "External id": 539426,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813779.371, "dur": 0.992,
    "args": {
      "External id": 539427,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813779.790, "dur": 0.507,
    "args": {
      "External id": 539428,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813781.240, "dur": 1.073,
    "args": {
      "External id": 539429,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813781.868, "dur": 0.379,
    "args": {
      "External id": 539430,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813782.776, "dur": 3.018,
    "args": {
      "External id": 539431,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813785.280, "dur": 0.448,
    "args": {
      "External id": 539432,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813786.089, "dur": 0.779,
    "args": {
      "External id": 539433,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813786.507, "dur": 0.294,
    "args": {
      "External id": 539434,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813787.118, "dur": 1.099,
    "args": {
      "External id": 539435,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813787.796, "dur": 0.352,
    "args": {
      "External id": 539436,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813788.476, "dur": 2.601,
    "args": {
      "External id": 539437,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813790.653, "dur": 0.355,
    "args": {
      "External id": 539438,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813791.335, "dur": 1.345,
    "args": {
      "External id": 539439,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813792.255, "dur": 0.354,
    "args": {
      "External id": 539440,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813792.928, "dur": 2.737,
    "args": {
      "External id": 539441,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813793.502, "dur": 1.859,
    "args": {
      "External id": 539442,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813795.943, "dur": 1.172,
    "args": {
      "External id": 539443,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813796.394, "dur": 0.655,
    "args": {
      "External id": 539444,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813797.366, "dur": 0.824,
    "args": {
      "External id": 539445,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813797.786, "dur": 0.339,
    "args": {
      "External id": 539446,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813798.446, "dur": 2.662,
    "args": {
      "External id": 539447,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 8998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813800.732, "dur": 0.310,
    "args": {
      "External id": 539448,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813801.378, "dur": 0.758,
    "args": {
      "External id": 539449,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813801.788, "dur": 0.281,
    "args": {
      "External id": 539450,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813802.417, "dur": 1.156,
    "args": {
      "External id": 539451,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813803.086, "dur": 0.420,
    "args": {
      "External id": 539452,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813803.917, "dur": 3.207,
    "args": {
      "External id": 539453,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813806.601, "dur": 0.455,
    "args": {
      "External id": 539454,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813807.385, "dur": 0.958,
    "args": {
      "External id": 539455,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813807.908, "dur": 0.371,
    "args": {
      "External id": 539456,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813808.590, "dur": 2.517,
    "args": {
      "External id": 539457,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813809.011, "dur": 1.939,
    "args": {
      "External id": 539458,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813811.373, "dur": 0.820,
    "args": {
      "External id": 539459,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813811.814, "dur": 0.312,
    "args": {
      "External id": 539460,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813812.562, "dur": 0.832,
    "args": {
      "External id": 539461,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813812.990, "dur": 0.336,
    "args": {
      "External id": 539462,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813813.734, "dur": 3.109,
    "args": {
      "External id": 539463,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813816.563, "dur": 0.213,
    "args": {
      "External id": 539464,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813817.126, "dur": 1.048,
    "args": {
      "External id": 539465,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813817.561, "dur": 0.547,
    "args": {
      "External id": 539466,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813818.432, "dur": 0.803,
    "args": {
      "External id": 539467,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813818.870, "dur": 0.298,
    "args": {
      "External id": 539468,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813819.489, "dur": 2.690,
    "args": {
      "External id": 539469,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813821.590, "dur": 0.524,
    "args": {
      "External id": 539470,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813822.441, "dur": 1.022,
    "args": {
      "External id": 539471,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813823.062, "dur": 0.334,
    "args": {
      "External id": 539472,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813823.713, "dur": 2.653,
    "args": {
      "External id": 539473,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813824.142, "dur": 2.071,
    "args": {
      "External id": 539474,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813826.633, "dur": 0.952,
    "args": {
      "External id": 539475,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813827.126, "dur": 0.393,
    "args": {
      "External id": 539476,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813828.047, "dur": 1.362,
    "args": {
      "External id": 539477,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813828.986, "dur": 0.357,
    "args": {
      "External id": 539478,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813829.661, "dur": 3.023,
    "args": {
      "External id": 539479,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813832.208, "dur": 0.407,
    "args": {
      "External id": 539480,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813832.958, "dur": 0.736,
    "args": {
      "External id": 539481,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813833.381, "dur": 0.246,
    "args": {
      "External id": 539482,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813833.947, "dur": 0.895,
    "args": {
      "External id": 539483,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813834.470, "dur": 0.305,
    "args": {
      "External id": 539484,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813835.091, "dur": 2.745,
    "args": {
      "External id": 539485,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813837.500, "dur": 0.268,
    "args": {
      "External id": 539486,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813838.096, "dur": 0.828,
    "args": {
      "External id": 539487,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813838.559, "dur": 0.296,
    "args": {
      "External id": 539488,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813839.192, "dur": 2.654,
    "args": {
      "External id": 539489,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813839.620, "dur": 2.151,
    "args": {
      "External id": 539490,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813842.096, "dur": 0.949,
    "args": {
      "External id": 539491,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813842.592, "dur": 0.380,
    "args": {
      "External id": 539492,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813843.320, "dur": 1.124,
    "args": {
      "External id": 539493,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813843.820, "dur": 0.555,
    "args": {
      "External id": 539494,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813844.884, "dur": 2.795,
    "args": {
      "External id": 539495,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813847.203, "dur": 0.407,
    "args": {
      "External id": 539496,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813847.947, "dur": 1.122,
    "args": {
      "External id": 539497,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813848.603, "dur": 0.399,
    "args": {
      "External id": 539498,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813849.319, "dur": 0.938,
    "args": {
      "External id": 539499,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813849.728, "dur": 0.454,
    "args": {
      "External id": 539500,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813850.520, "dur": 3.288,
    "args": {
      "External id": 539501,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813853.419, "dur": 0.323,
    "args": {
      "External id": 539502,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813854.070, "dur": 0.934,
    "args": {
      "External id": 539503,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813854.526, "dur": 0.409,
    "args": {
      "External id": 539504,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813855.252, "dur": 2.899,
    "args": {
      "External id": 539505,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813855.728, "dur": 2.162,
    "args": {
      "External id": 539506,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813858.412, "dur": 1.107,
    "args": {
      "External id": 539507,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813859.053, "dur": 0.398,
    "args": {
      "External id": 539508,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813859.947, "dur": 0.847,
    "args": {
      "External id": 539509,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813860.377, "dur": 0.349,
    "args": {
      "External id": 539510,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813861.047, "dur": 3.074,
    "args": {
      "External id": 539511,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813863.560, "dur": 0.486,
    "args": {
      "External id": 539512,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813864.399, "dur": 1.154,
    "args": {
      "External id": 539513,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813865.228, "dur": 0.256,
    "args": {
      "External id": 539514,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813865.809, "dur": 0.770,
    "args": {
      "External id": 539515,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813866.229, "dur": 0.283,
    "args": {
      "External id": 539516,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813866.829, "dur": 2.525,
    "args": {
      "External id": 539517,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813868.690, "dur": 0.602,
    "args": {
      "External id": 539518,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813869.697, "dur": 0.982,
    "args": {
      "External id": 539519,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813870.253, "dur": 0.358,
    "args": {
      "External id": 539520,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813870.939, "dur": 2.634,
    "args": {
      "External id": 539521,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813871.347, "dur": 2.065,
    "args": {
      "External id": 539522,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813873.831, "dur": 0.903,
    "args": {
      "External id": 539523,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813874.340, "dur": 0.327,
    "args": {
      "External id": 539524,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813875.320, "dur": 0.771,
    "args": {
      "External id": 539525,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813875.756, "dur": 0.267,
    "args": {
      "External id": 539526,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813876.526, "dur": 3.018,
    "args": {
      "External id": 539527,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813879.094, "dur": 0.382,
    "args": {
      "External id": 539528,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813879.837, "dur": 0.942,
    "args": {
      "External id": 539529,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813880.263, "dur": 0.447,
    "args": {
      "External id": 539530,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813881.035, "dur": 1.064,
    "args": {
      "External id": 539531,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813881.626, "dur": 0.404,
    "args": {
      "External id": 539532,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813882.368, "dur": 3.764,
    "args": {
      "External id": 539533,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813885.706, "dur": 0.359,
    "args": {
      "External id": 539534,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813886.387, "dur": 0.999,
    "args": {
      "External id": 539535,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813886.887, "dur": 0.432,
    "args": {
      "External id": 539536,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813887.631, "dur": 2.482,
    "args": {
      "External id": 539537,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813888.044, "dur": 1.814,
    "args": {
      "External id": 539538,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813890.368, "dur": 0.779,
    "args": {
      "External id": 539539,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813890.832, "dur": 0.245,
    "args": {
      "External id": 539540,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813891.405, "dur": 1.011,
    "args": {
      "External id": 539541,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813891.823, "dur": 0.524,
    "args": {
      "External id": 539542,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813892.751, "dur": 3.005,
    "args": {
      "External id": 539543,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813895.295, "dur": 0.394,
    "args": {
      "External id": 539544,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813896.025, "dur": 0.856,
    "args": {
      "External id": 539545,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813896.423, "dur": 0.382,
    "args": {
      "External id": 539546,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813897.131, "dur": 0.883,
    "args": {
      "External id": 539547,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813897.547, "dur": 0.398,
    "args": {
      "External id": 539548,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813898.265, "dur": 3.227,
    "args": {
      "External id": 539549,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813901.065, "dur": 0.360,
    "args": {
      "External id": 539550,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813901.761, "dur": 1.243,
    "args": {
      "External id": 539551,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813902.325, "dur": 0.609,
    "args": {
      "External id": 539552,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813903.252, "dur": 2.413,
    "args": {
      "External id": 539553,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813903.666, "dur": 1.929,
    "args": {
      "External id": 539554,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813905.917, "dur": 0.912,
    "args": {
      "External id": 539555,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813906.358, "dur": 0.405,
    "args": {
      "External id": 539556,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813907.086, "dur": 0.919,
    "args": {
      "External id": 539557,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813907.565, "dur": 0.373,
    "args": {
      "External id": 539558,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813908.277, "dur": 2.832,
    "args": {
      "External id": 539559,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813910.725, "dur": 0.309,
    "args": {
      "External id": 539560,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813911.382, "dur": 0.801,
    "args": {
      "External id": 539561,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813911.858, "dur": 0.260,
    "args": {
      "External id": 539562,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813912.581, "dur": 0.929,
    "args": {
      "External id": 539563,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813913.086, "dur": 0.356,
    "args": {
      "External id": 539564,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813913.758, "dur": 2.652,
    "args": {
      "External id": 539565,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813915.923, "dur": 0.420,
    "args": {
      "External id": 539566,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813916.678, "dur": 0.887,
    "args": {
      "External id": 539567,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813917.115, "dur": 0.385,
    "args": {
      "External id": 539568,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813917.812, "dur": 2.429,
    "args": {
      "External id": 539569,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813918.246, "dur": 1.921,
    "args": {
      "External id": 539570,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813920.489, "dur": 1.017,
    "args": {
      "External id": 539571,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813920.994, "dur": 0.443,
    "args": {
      "External id": 539572,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813921.943, "dur": 1.053,
    "args": {
      "External id": 539573,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813922.437, "dur": 0.489,
    "args": {
      "External id": 539574,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813923.334, "dur": 2.651,
    "args": {
      "External id": 539575,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813925.661, "dur": 0.256,
    "args": {
      "External id": 539576,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813926.275, "dur": 0.766,
    "args": {
      "External id": 539577,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813926.705, "dur": 0.266,
    "args": {
      "External id": 539578,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813927.287, "dur": 0.891,
    "args": {
      "External id": 539579,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813927.703, "dur": 0.408,
    "args": {
      "External id": 539580,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813928.437, "dur": 2.574,
    "args": {
      "External id": 539581,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813930.611, "dur": 0.334,
    "args": {
      "External id": 539582,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813931.275, "dur": 1.215,
    "args": {
      "External id": 539583,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813932.036, "dur": 0.387,
    "args": {
      "External id": 539584,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813932.742, "dur": 2.576,
    "args": {
      "External id": 539585,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813933.233, "dur": 2.012,
    "args": {
      "External id": 539586,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813935.616, "dur": 1.102,
    "args": {
      "External id": 539587,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813936.277, "dur": 0.368,
    "args": {
      "External id": 539588,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813936.971, "dur": 1.100,
    "args": {
      "External id": 539589,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813937.453, "dur": 0.551,
    "args": {
      "External id": 539590,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813938.335, "dur": 2.795,
    "args": {
      "External id": 539591,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813940.579, "dur": 0.481,
    "args": {
      "External id": 539592,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813941.405, "dur": 1.051,
    "args": {
      "External id": 539593,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813941.819, "dur": 0.564,
    "args": {
      "External id": 539594,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813942.707, "dur": 0.899,
    "args": {
      "External id": 539595,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813943.247, "dur": 0.289,
    "args": {
      "External id": 539596,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813943.952, "dur": 2.768,
    "args": {
      "External id": 539597,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813946.395, "dur": 0.258,
    "args": {
      "External id": 539598,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813947.005, "dur": 0.870,
    "args": {
      "External id": 539599,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813947.515, "dur": 0.294,
    "args": {
      "External id": 539600,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813948.329, "dur": 2.649,
    "args": {
      "External id": 539601,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813948.788, "dur": 2.031,
    "args": {
      "External id": 539602,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813951.233, "dur": 1.132,
    "args": {
      "External id": 539603,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813951.678, "dur": 0.620,
    "args": {
      "External id": 539604,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813952.704, "dur": 0.763,
    "args": {
      "External id": 539605,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813953.126, "dur": 0.276,
    "args": {
      "External id": 539606,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813953.966, "dur": 3.598,
    "args": {
      "External id": 539607,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813956.931, "dur": 0.565,
    "args": {
      "External id": 539608,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813957.836, "dur": 1.102,
    "args": {
      "External id": 539609,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813958.350, "dur": 0.432,
    "args": {
      "External id": 539610,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813959.196, "dur": 1.085,
    "args": {
      "External id": 539611,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813959.683, "dur": 0.533,
    "args": {
      "External id": 539612,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813960.563, "dur": 2.239,
    "args": {
      "External id": 539613,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813962.503, "dur": 0.228,
    "args": {
      "External id": 539614,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813963.064, "dur": 0.832,
    "args": {
      "External id": 539615,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813963.500, "dur": 0.327,
    "args": {
      "External id": 539616,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813964.170, "dur": 2.434,
    "args": {
      "External id": 539617,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813964.704, "dur": 1.830,
    "args": {
      "External id": 539618,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813966.855, "dur": 0.769,
    "args": {
      "External id": 539619,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813967.270, "dur": 0.286,
    "args": {
      "External id": 539620,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813968.202, "dur": 1.042,
    "args": {
      "External id": 539621,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813968.634, "dur": 0.542,
    "args": {
      "External id": 539622,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813969.611, "dur": 2.414,
    "args": {
      "External id": 539623,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813971.682, "dur": 0.278,
    "args": {
      "External id": 539624,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813972.297, "dur": 0.792,
    "args": {
      "External id": 539625,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813972.781, "dur": 0.242,
    "args": {
      "External id": 539626,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813973.449, "dur": 0.902,
    "args": {
      "External id": 539627,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813973.884, "dur": 0.401,
    "args": {
      "External id": 539628,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813974.796, "dur": 2.554,
    "args": {
      "External id": 539629,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813977.061, "dur": 0.224,
    "args": {
      "External id": 539630,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813977.771, "dur": 0.912,
    "args": {
      "External id": 539631,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813978.258, "dur": 0.358,
    "args": {
      "External id": 539632,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813978.978, "dur": 2.699,
    "args": {
      "External id": 539633,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813979.418, "dur": 2.188,
    "args": {
      "External id": 539634,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813981.934, "dur": 0.930,
    "args": {
      "External id": 539635,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813982.426, "dur": 0.372,
    "args": {
      "External id": 539636,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813983.419, "dur": 0.979,
    "args": {
      "External id": 539637,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813983.847, "dur": 0.483,
    "args": {
      "External id": 539638,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813984.664, "dur": 2.599,
    "args": {
      "External id": 539639,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813986.850, "dur": 0.346,
    "args": {
      "External id": 539640,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813987.547, "dur": 0.803,
    "args": {
      "External id": 539641,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813987.988, "dur": 0.297,
    "args": {
      "External id": 539642,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813988.714, "dur": 0.785,
    "args": {
      "External id": 539643,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813989.143, "dur": 0.291,
    "args": {
      "External id": 539644,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813990.018, "dur": 2.820,
    "args": {
      "External id": 539645,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813992.381, "dur": 0.389,
    "args": {
      "External id": 539646,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622813993.185, "dur": 0.985,
    "args": {
      "External id": 539647,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622813993.673, "dur": 0.432,
    "args": {
      "External id": 539648,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814010.984, "dur": 3.283,
    "args": {
      "External id": 539649,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814011.865, "dur": 2.324,
    "args": {
      "External id": 539650,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814014.609, "dur": 0.929,
    "args": {
      "External id": 539651,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814015.255, "dur": 0.215,
    "args": {
      "External id": 539652,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814015.830, "dur": 0.932,
    "args": {
      "External id": 539653,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814016.249, "dur": 0.445,
    "args": {
      "External id": 539654,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814017.045, "dur": 2.780,
    "args": {
      "External id": 539655,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814019.416, "dur": 0.340,
    "args": {
      "External id": 539656,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814020.174, "dur": 0.768,
    "args": {
      "External id": 539657,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814020.596, "dur": 0.273,
    "args": {
      "External id": 539658,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814021.225, "dur": 0.922,
    "args": {
      "External id": 539659,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814021.766, "dur": 0.312,
    "args": {
      "External id": 539660,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814022.543, "dur": 2.902,
    "args": {
      "External id": 539661,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814025.009, "dur": 0.367,
    "args": {
      "External id": 539662,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814025.705, "dur": 0.880,
    "args": {
      "External id": 539663,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814026.261, "dur": 0.255,
    "args": {
      "External id": 539664,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814026.845, "dur": 2.825,
    "args": {
      "External id": 539665,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814027.353, "dur": 2.246,
    "args": {
      "External id": 539666,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814029.985, "dur": 1.008,
    "args": {
      "External id": 539667,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814030.490, "dur": 0.434,
    "args": {
      "External id": 539668,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814031.548, "dur": 0.930,
    "args": {
      "External id": 539669,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814032.002, "dur": 0.408,
    "args": {
      "External id": 539670,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814032.762, "dur": 2.479,
    "args": {
      "External id": 539671,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814034.896, "dur": 0.275,
    "args": {
      "External id": 539672,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814035.540, "dur": 0.825,
    "args": {
      "External id": 539673,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814036.029, "dur": 0.269,
    "args": {
      "External id": 539674,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814036.718, "dur": 10.661,
    "args": {
      "External id": 539675,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814041.095, "dur": 6.097,
    "args": {
      "External id": 539676,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814047.664, "dur": 1.063,
    "args": {
      "External id": 539677,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814048.205, "dur": 0.445,
    "args": {
      "External id": 539678,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814049.000, "dur": 0.969,
    "args": {
      "External id": 539679,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814049.451, "dur": 0.444,
    "args": {
      "External id": 539680,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814050.473, "dur": 2.812,
    "args": {
      "External id": 539681,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814052.861, "dur": 0.353,
    "args": {
      "External id": 539682,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814053.584, "dur": 0.842,
    "args": {
      "External id": 539683,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814054.032, "dur": 0.322,
    "args": {
      "External id": 539684,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814054.851, "dur": 1.081,
    "args": {
      "External id": 539685,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814055.489, "dur": 0.369,
    "args": {
      "External id": 539686,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814056.234, "dur": 2.441,
    "args": {
      "External id": 539687,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814058.288, "dur": 0.316,
    "args": {
      "External id": 539688,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814059.032, "dur": 0.799,
    "args": {
      "External id": 539689,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814059.475, "dur": 0.285,
    "args": {
      "External id": 539690,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814060.096, "dur": 2.975,
    "args": {
      "External id": 539691,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814060.747, "dur": 2.165,
    "args": {
      "External id": 539692,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814063.551, "dur": 1.369,
    "args": {
      "External id": 539693,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814064.189, "dur": 0.664,
    "args": {
      "External id": 539694,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814065.602, "dur": 0.828,
    "args": {
      "External id": 539695,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814066.072, "dur": 0.289,
    "args": {
      "External id": 539696,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814066.701, "dur": 2.924,
    "args": {
      "External id": 539697,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814069.314, "dur": 0.240,
    "args": {
      "External id": 539698,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 4183438, "tid": 4183438,
    "ts": 679622814069.920, "dur": 0.947,
    "args": {
      "External id": 539699,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814070.374, "dur": 0.420,
    "args": {
      "External id": 539700,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 4183438, "tid": 4183438,
    "ts": 679622814089.579, "dur": 127.159,
    "args": {
      "External id": 539701,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linalg_vector_norm", "pid": 4183438, "tid": 4183438,
    "ts": 679622814306.273, "dur": 129.776,
    "args": {
      "External id": 539702,"Record function id": 0, "Concrete Inputs": ["", "2.", "", "False", ""], "Input type": ["float", "Scalar", "", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[219], [], [], [], []], "Ev Idx": 9253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linalg_vector_norm", "pid": 4183438, "tid": 4183438,
    "ts": 679622814368.528, "dur": 48.845,
    "args": {
      "External id": 539703,"Record function id": 0, "Concrete Inputs": ["", "2.", "", "False", ""], "Input type": ["float", "Scalar", "", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[219], [], [], [], []], "Ev Idx": 9254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622814381.271, "dur": 0.989,
    "args": {
      "External id": 539704,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Redistribute", "pid": 4183438, "tid": 4183438,
    "ts": 679622814746.136, "dur": 863.059,
    "args": {
      "External id": 539705,"Sequence number": 6319636, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "False"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 4183438, "tid": 4183438,
    "ts": 679622814799.071, "dur": 49.402,
    "args": {
      "External id": 539706,"Record function id": 0, "Concrete Inputs": ["", "2."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622814804.011, "dur": 1.455,
    "args": {
      "External id": 539707,"Record function id": 0, "Concrete Inputs": ["", "2."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622814807.274, "dur": 0.543,
    "args": {
      "External id": 539708,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622814872.552, "dur": 436.864,
    "args": {
      "External id": 539709,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622814876.961, "dur": 44.649,
    "args": {
      "External id": 539710,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622814879.840, "dur": 8.808,
    "args": {
      "External id": 539711,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622814883.719, "dur": 4.185,
    "args": {
      "External id": 539712,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622814890.155, "dur": 30.887,
    "args": {
      "External id": 539713,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 4183438, "tid": 4183438,
    "ts": 679622814930.735, "dur": 374.650,
    "args": {
      "External id": 539714,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[]], [], [], [], []], "Input Dims": [[[]], [], [], [], []], "Ev Idx": 9265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622814957.155, "dur": 341.562,
    "args": {
      "External id": 539715,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[[]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 9266, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622814974.174, "dur": 318.393,
    "args": {
      "External id": 539716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 4183438, "tid": 4183438,
    "ts": 679622815372.832, "dur": 198.897,
    "args": {
      "External id": 539717,"Record function id": 0, "Concrete Inputs": ["", "0.5"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622815465.480, "dur": 27.346,
    "args": {
      "External id": 539718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622815480.216, "dur": 4.688,
    "args": {
      "External id": 539719,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 9270, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 4183438, "tid": 4183438,
    "ts": 679622815521.150, "dur": 44.422,
    "args": {
      "External id": 539720,"Record function id": 0, "Concrete Inputs": ["", "0.5"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622815523.618, "dur": 1.174,
    "args": {
      "External id": 539721,"Record function id": 0, "Concrete Inputs": ["", "0.5"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622815526.111, "dur": 0.620,
    "args": {
      "External id": 539722,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_ToTorchTensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622815625.485, "dur": 22.178,
    "args": {
      "External id": 539723,"Sequence number": 6319637, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 4183438, "tid": 4183438,
    "ts": 679622815636.288, "dur": 8.294,
    "args": {
      "External id": 539724,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622815638.893, "dur": 5.472,
    "args": {
      "External id": 539725,"Record function id": 0, "Concrete Inputs": ["", "[]"], "Input type": ["float", "ScalarList"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 4183438, "tid": 4183438,
    "ts": 679622816047.437, "dur": 40.935,
    "args": {
      "External id": 539726,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "double", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 4183438, "tid": 4183438,
    "ts": 679622816098.580, "dur": 41.986,
    "args": {
      "External id": 539727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 4183438, "tid": 4183438,
    "ts": 679622816149.840, "dur": 31.951,
    "args": {
      "External id": 539728,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 4183438, "tid": 4183438,
    "ts": 679622816199.259, "dur": 27.587,
    "args": {
      "External id": 539729,"Record function id": 0, "Concrete Inputs": ["", "", "1."], "Input type": ["float", "", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622816203.795, "dur": 0.462,
    "args": {
      "External id": 539730,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622816244.306, "dur": 0.446,
    "args": {
      "External id": 539731,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 9282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_mul_", "pid": 4183438, "tid": 4183438,
    "ts": 679622816331.557, "dur": 514.459,
    "args": {
      "External id": 539732,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["TensorList", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_mul_", "pid": 4183438, "tid": 4183438,
    "ts": 679622816758.638, "dur": 60.657,
    "args": {
      "External id": 539733,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["TensorList", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::isnan", "pid": 4183438, "tid": 4183438,
    "ts": 679622816884.659, "dur": 25.487,
    "args": {
      "External id": 539734,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 4183438, "tid": 4183438,
    "ts": 679622816888.269, "dur": 21.326,
    "args": {
      "External id": 539735,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622816913.753, "dur": 347.796,
    "args": {
      "External id": 539736,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622816915.379, "dur": 345.956,
    "args": {
      "External id": 539737,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622816916.946, "dur": 343.669,
    "args": {
      "External id": 539738,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::isinf", "pid": 4183438, "tid": 4183438,
    "ts": 679622817267.490, "dur": 60.756,
    "args": {
      "External id": 539739,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622817270.931, "dur": 33.636,
    "args": {
      "External id": 539740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622817277.398, "dur": 3.121,
    "args": {
      "External id": 539741,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 4183438, "tid": 4183438,
    "ts": 679622817282.109, "dur": 22.165,
    "args": {
      "External id": 539742,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], [1]], "Input Dims": [[], [0]], "Ev Idx": 9293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 4183438, "tid": 4183438,
    "ts": 679622817286.983, "dur": 2.561,
    "args": {
      "External id": 539743,"Record function id": 0, "Concrete Inputs": ["", "[]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622817306.294, "dur": 21.111,
    "args": {
      "External id": 539744,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622817330.458, "dur": 35.313,
    "args": {
      "External id": 539745,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622817331.224, "dur": 34.404,
    "args": {
      "External id": 539746,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622817332.266, "dur": 33.078,
    "args": {
      "External id": 539747,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9298
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.step#OptimizersContainer.step", "pid": 4183438, "tid": 4183438,
    "ts": 679622817394.207, "dur": 4368.061,
    "args": {
      "External id": 539748,"Record function id": 0, "Ev Idx": 9299
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.step#AdamW.step", "pid": 4183438, "tid": 4183438,
    "ts": 679622817424.168, "dur": 4316.792,
    "args": {
      "External id": 539749,"Record function id": 0, "Ev Idx": 9300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_add_", "pid": 4183438, "tid": 4183438,
    "ts": 679622818415.275, "dur": 195.879,
    "args": {
      "External id": 539750,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818432.737, "dur": 0.858,
    "args": {
      "External id": 539751,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818434.388, "dur": 0.195,
    "args": {
      "External id": 539752,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818435.216, "dur": 0.172,
    "args": {
      "External id": 539753,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818435.791, "dur": 0.217,
    "args": {
      "External id": 539754,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818436.316, "dur": 0.063,
    "args": {
      "External id": 539755,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818436.817, "dur": 0.201,
    "args": {
      "External id": 539756,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818437.547, "dur": 0.183,
    "args": {
      "External id": 539757,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818438.121, "dur": 0.203,
    "args": {
      "External id": 539758,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818438.628, "dur": 0.223,
    "args": {
      "External id": 539759,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818439.293, "dur": 0.177,
    "args": {
      "External id": 539760,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818439.843, "dur": 0.065,
    "args": {
      "External id": 539761,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818440.369, "dur": 0.059,
    "args": {
      "External id": 539762,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818440.809, "dur": 0.063,
    "args": {
      "External id": 539763,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818441.317, "dur": 0.057,
    "args": {
      "External id": 539764,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818441.753, "dur": 0.069,
    "args": {
      "External id": 539765,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818442.201, "dur": 0.054,
    "args": {
      "External id": 539766,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818442.607, "dur": 0.070,
    "args": {
      "External id": 539767,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818443.021, "dur": 0.070,
    "args": {
      "External id": 539768,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818443.375, "dur": 0.068,
    "args": {
      "External id": 539769,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818443.814, "dur": 0.071,
    "args": {
      "External id": 539770,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818444.188, "dur": 0.069,
    "args": {
      "External id": 539771,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818444.755, "dur": 0.069,
    "args": {
      "External id": 539772,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818445.347, "dur": 0.068,
    "args": {
      "External id": 539773,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818445.713, "dur": 0.067,
    "args": {
      "External id": 539774,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818446.034, "dur": 0.102,
    "args": {
      "External id": 539775,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818446.415, "dur": 0.104,
    "args": {
      "External id": 539776,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818446.913, "dur": 0.078,
    "args": {
      "External id": 539777,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818447.364, "dur": 0.280,
    "args": {
      "External id": 539778,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818448.021, "dur": 0.069,
    "args": {
      "External id": 539779,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818448.484, "dur": 0.101,
    "args": {
      "External id": 539780,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818449.014, "dur": 0.187,
    "args": {
      "External id": 539781,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818449.560, "dur": 0.071,
    "args": {
      "External id": 539782,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818450.161, "dur": 0.092,
    "args": {
      "External id": 539783,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818450.575, "dur": 0.087,
    "args": {
      "External id": 539784,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818451.059, "dur": 0.211,
    "args": {
      "External id": 539785,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818451.559, "dur": 0.202,
    "args": {
      "External id": 539786,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818452.264, "dur": 0.057,
    "args": {
      "External id": 539787,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818452.730, "dur": 0.051,
    "args": {
      "External id": 539788,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818453.294, "dur": 0.067,
    "args": {
      "External id": 539789,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818453.630, "dur": 0.186,
    "args": {
      "External id": 539790,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818454.147, "dur": 0.214,
    "args": {
      "External id": 539791,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818454.680, "dur": 0.197,
    "args": {
      "External id": 539792,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818455.237, "dur": 0.100,
    "args": {
      "External id": 539793,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818455.804, "dur": 0.102,
    "args": {
      "External id": 539794,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818456.276, "dur": 0.065,
    "args": {
      "External id": 539795,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818456.727, "dur": 0.078,
    "args": {
      "External id": 539796,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818457.181, "dur": 0.186,
    "args": {
      "External id": 539797,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818457.748, "dur": 0.079,
    "args": {
      "External id": 539798,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818458.193, "dur": 0.102,
    "args": {
      "External id": 539799,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818458.667, "dur": 0.076,
    "args": {
      "External id": 539800,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818459.119, "dur": 0.065,
    "args": {
      "External id": 539801,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818459.555, "dur": 0.211,
    "args": {
      "External id": 539802,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818460.126, "dur": 0.062,
    "args": {
      "External id": 539803,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818460.551, "dur": 0.066,
    "args": {
      "External id": 539804,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818461.021, "dur": 0.063,
    "args": {
      "External id": 539805,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818461.420, "dur": 0.076,
    "args": {
      "External id": 539806,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818461.840, "dur": 0.068,
    "args": {
      "External id": 539807,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818462.484, "dur": 0.069,
    "args": {
      "External id": 539808,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818462.826, "dur": 0.062,
    "args": {
      "External id": 539809,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818463.455, "dur": 0.067,
    "args": {
      "External id": 539810,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818463.921, "dur": 0.063,
    "args": {
      "External id": 539811,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818464.333, "dur": 0.066,
    "args": {
      "External id": 539812,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818464.775, "dur": 0.062,
    "args": {
      "External id": 539813,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818465.147, "dur": 0.068,
    "args": {
      "External id": 539814,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818465.705, "dur": 0.056,
    "args": {
      "External id": 539815,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818466.140, "dur": 0.067,
    "args": {
      "External id": 539816,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818466.573, "dur": 0.088,
    "args": {
      "External id": 539817,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818467.048, "dur": 0.099,
    "args": {
      "External id": 539818,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818467.503, "dur": 0.049,
    "args": {
      "External id": 539819,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818467.841, "dur": 0.199,
    "args": {
      "External id": 539820,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818468.399, "dur": 0.076,
    "args": {
      "External id": 539821,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818468.998, "dur": 0.050,
    "args": {
      "External id": 539822,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818469.423, "dur": 0.104,
    "args": {
      "External id": 539823,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818469.895, "dur": 0.099,
    "args": {
      "External id": 539824,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818470.355, "dur": 0.248,
    "args": {
      "External id": 539825,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818470.982, "dur": 0.093,
    "args": {
      "External id": 539826,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818471.401, "dur": 0.066,
    "args": {
      "External id": 539827,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818471.936, "dur": 0.179,
    "args": {
      "External id": 539828,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818472.511, "dur": 0.052,
    "args": {
      "External id": 539829,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818472.915, "dur": 0.067,
    "args": {
      "External id": 539830,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818473.348, "dur": 0.066,
    "args": {
      "External id": 539831,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818473.771, "dur": 0.070,
    "args": {
      "External id": 539832,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818474.290, "dur": 0.067,
    "args": {
      "External id": 539833,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818474.739, "dur": 0.068,
    "args": {
      "External id": 539834,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818475.194, "dur": 0.067,
    "args": {
      "External id": 539835,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818475.700, "dur": 0.066,
    "args": {
      "External id": 539836,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818476.177, "dur": 0.066,
    "args": {
      "External id": 539837,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818476.637, "dur": 0.067,
    "args": {
      "External id": 539838,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818477.039, "dur": 0.058,
    "args": {
      "External id": 539839,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818477.452, "dur": 0.062,
    "args": {
      "External id": 539840,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818477.964, "dur": 0.060,
    "args": {
      "External id": 539841,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818478.310, "dur": 0.058,
    "args": {
      "External id": 539842,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818478.759, "dur": 0.054,
    "args": {
      "External id": 539843,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818479.188, "dur": 0.060,
    "args": {
      "External id": 539844,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818479.628, "dur": 0.060,
    "args": {
      "External id": 539845,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818479.965, "dur": 0.061,
    "args": {
      "External id": 539846,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818480.792, "dur": 0.062,
    "args": {
      "External id": 539847,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818481.197, "dur": 0.063,
    "args": {
      "External id": 539848,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818481.621, "dur": 0.170,
    "args": {
      "External id": 539849,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818482.148, "dur": 0.056,
    "args": {
      "External id": 539850,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818482.517, "dur": 0.090,
    "args": {
      "External id": 539851,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818483.351, "dur": 0.090,
    "args": {
      "External id": 539852,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818483.797, "dur": 0.087,
    "args": {
      "External id": 539853,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818484.343, "dur": 0.190,
    "args": {
      "External id": 539854,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818484.879, "dur": 0.105,
    "args": {
      "External id": 539855,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818485.264, "dur": 0.200,
    "args": {
      "External id": 539856,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818485.870, "dur": 0.204,
    "args": {
      "External id": 539857,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818486.446, "dur": 0.097,
    "args": {
      "External id": 539858,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818486.920, "dur": 0.068,
    "args": {
      "External id": 539859,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818487.398, "dur": 0.082,
    "args": {
      "External id": 539860,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818487.773, "dur": 0.082,
    "args": {
      "External id": 539861,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818488.177, "dur": 0.200,
    "args": {
      "External id": 539862,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818488.747, "dur": 0.097,
    "args": {
      "External id": 539863,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818489.097, "dur": 0.094,
    "args": {
      "External id": 539864,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818489.500, "dur": 0.220,
    "args": {
      "External id": 539865,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818489.997, "dur": 0.065,
    "args": {
      "External id": 539866,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818490.432, "dur": 0.059,
    "args": {
      "External id": 539867,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818490.781, "dur": 0.091,
    "args": {
      "External id": 539868,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818491.222, "dur": 0.069,
    "args": {
      "External id": 539869,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818491.621, "dur": 0.073,
    "args": {
      "External id": 539870,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818492.184, "dur": 0.071,
    "args": {
      "External id": 539871,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818492.611, "dur": 0.069,
    "args": {
      "External id": 539872,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818493.010, "dur": 0.069,
    "args": {
      "External id": 539873,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818493.473, "dur": 0.066,
    "args": {
      "External id": 539874,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818493.937, "dur": 0.049,
    "args": {
      "External id": 539875,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818494.371, "dur": 0.067,
    "args": {
      "External id": 539876,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818494.713, "dur": 0.064,
    "args": {
      "External id": 539877,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818495.074, "dur": 0.074,
    "args": {
      "External id": 539878,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818495.506, "dur": 0.061,
    "args": {
      "External id": 539879,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818495.926, "dur": 0.059,
    "args": {
      "External id": 539880,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818496.346, "dur": 0.067,
    "args": {
      "External id": 539881,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818496.755, "dur": 0.071,
    "args": {
      "External id": 539882,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818497.156, "dur": 0.064,
    "args": {
      "External id": 539883,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818497.518, "dur": 0.064,
    "args": {
      "External id": 539884,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818497.955, "dur": 0.066,
    "args": {
      "External id": 539885,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818498.356, "dur": 0.068,
    "args": {
      "External id": 539886,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818498.734, "dur": 0.068,
    "args": {
      "External id": 539887,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818499.188, "dur": 0.067,
    "args": {
      "External id": 539888,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818499.736, "dur": 0.068,
    "args": {
      "External id": 539889,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818500.117, "dur": 0.089,
    "args": {
      "External id": 539890,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818500.585, "dur": 0.070,
    "args": {
      "External id": 539891,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818501.111, "dur": 0.065,
    "args": {
      "External id": 539892,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818501.551, "dur": 0.066,
    "args": {
      "External id": 539893,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818501.866, "dur": 0.211,
    "args": {
      "External id": 539894,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818502.333, "dur": 0.108,
    "args": {
      "External id": 539895,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818502.694, "dur": 0.259,
    "args": {
      "External id": 539896,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818503.205, "dur": 0.086,
    "args": {
      "External id": 539897,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818503.829, "dur": 0.086,
    "args": {
      "External id": 539898,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818504.188, "dur": 0.056,
    "args": {
      "External id": 539899,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818504.804, "dur": 0.201,
    "args": {
      "External id": 539900,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818505.351, "dur": 0.068,
    "args": {
      "External id": 539901,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818506.093, "dur": 0.082,
    "args": {
      "External id": 539902,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818506.417, "dur": 0.049,
    "args": {
      "External id": 539903,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818507.065, "dur": 0.088,
    "args": {
      "External id": 539904,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818507.393, "dur": 0.059,
    "args": {
      "External id": 539905,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818508.210, "dur": 0.055,
    "args": {
      "External id": 539906,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818508.535, "dur": 0.052,
    "args": {
      "External id": 539907,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818509.308, "dur": 0.056,
    "args": {
      "External id": 539908,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818509.602, "dur": 0.053,
    "args": {
      "External id": 539909,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818510.237, "dur": 0.068,
    "args": {
      "External id": 539910,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818510.541, "dur": 0.049,
    "args": {
      "External id": 539911,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818511.471, "dur": 0.087,
    "args": {
      "External id": 539912,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818511.800, "dur": 0.056,
    "args": {
      "External id": 539913,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818512.362, "dur": 0.068,
    "args": {
      "External id": 539914,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818512.673, "dur": 0.051,
    "args": {
      "External id": 539915,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818513.199, "dur": 0.065,
    "args": {
      "External id": 539916,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818513.525, "dur": 0.057,
    "args": {
      "External id": 539917,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818514.381, "dur": 0.066,
    "args": {
      "External id": 539918,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818514.691, "dur": 0.055,
    "args": {
      "External id": 539919,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818515.249, "dur": 0.067,
    "args": {
      "External id": 539920,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818515.558, "dur": 0.053,
    "args": {
      "External id": 539921,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818516.155, "dur": 0.064,
    "args": {
      "External id": 539922,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818516.463, "dur": 0.052,
    "args": {
      "External id": 539923,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818517.269, "dur": 0.065,
    "args": {
      "External id": 539924,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818517.577, "dur": 0.049,
    "args": {
      "External id": 539925,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818518.366, "dur": 0.187,
    "args": {
      "External id": 539926,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818518.799, "dur": 0.048,
    "args": {
      "External id": 539927,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818519.577, "dur": 0.161,
    "args": {
      "External id": 539928,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818519.982, "dur": 0.079,
    "args": {
      "External id": 539929,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818520.856, "dur": 0.084,
    "args": {
      "External id": 539930,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818521.188, "dur": 0.066,
    "args": {
      "External id": 539931,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818521.999, "dur": 0.066,
    "args": {
      "External id": 539932,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818522.431, "dur": 0.064,
    "args": {
      "External id": 539933,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818523.197, "dur": 0.089,
    "args": {
      "External id": 539934,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818523.637, "dur": 0.069,
    "args": {
      "External id": 539935,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818524.437, "dur": 0.065,
    "args": {
      "External id": 539936,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818524.864, "dur": 0.068,
    "args": {
      "External id": 539937,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818525.564, "dur": 0.068,
    "args": {
      "External id": 539938,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818525.992, "dur": 0.067,
    "args": {
      "External id": 539939,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818526.444, "dur": 0.067,
    "args": {
      "External id": 539940,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818526.879, "dur": 0.067,
    "args": {
      "External id": 539941,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818527.424, "dur": 0.063,
    "args": {
      "External id": 539942,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818527.732, "dur": 0.059,
    "args": {
      "External id": 539943,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818528.354, "dur": 0.067,
    "args": {
      "External id": 539944,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818528.659, "dur": 0.052,
    "args": {
      "External id": 539945,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818529.539, "dur": 0.067,
    "args": {
      "External id": 539946,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818529.846, "dur": 0.059,
    "args": {
      "External id": 539947,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818530.596, "dur": 0.069,
    "args": {
      "External id": 539948,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818530.911, "dur": 0.050,
    "args": {
      "External id": 539949,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818531.412, "dur": 0.089,
    "args": {
      "External id": 539950,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818531.745, "dur": 0.055,
    "args": {
      "External id": 539951,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818532.438, "dur": 0.066,
    "args": {
      "External id": 539952,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818532.751, "dur": 0.050,
    "args": {
      "External id": 539953,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818533.391, "dur": 0.066,
    "args": {
      "External id": 539954,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818533.699, "dur": 0.053,
    "args": {
      "External id": 539955,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818534.211, "dur": 0.063,
    "args": {
      "External id": 539956,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818534.538, "dur": 0.055,
    "args": {
      "External id": 539957,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818535.102, "dur": 0.066,
    "args": {
      "External id": 539958,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818535.415, "dur": 0.056,
    "args": {
      "External id": 539959,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818536.059, "dur": 0.067,
    "args": {
      "External id": 539960,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818536.391, "dur": 0.055,
    "args": {
      "External id": 539961,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818537.137, "dur": 0.064,
    "args": {
      "External id": 539962,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818537.444, "dur": 0.051,
    "args": {
      "External id": 539963,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818537.940, "dur": 0.067,
    "args": {
      "External id": 539964,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818538.251, "dur": 0.052,
    "args": {
      "External id": 539965,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818538.925, "dur": 0.069,
    "args": {
      "External id": 539966,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818539.237, "dur": 0.052,
    "args": {
      "External id": 539967,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818539.692, "dur": 0.067,
    "args": {
      "External id": 539968,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 4183438, "tid": 4183438,
    "ts": 679622818540.006, "dur": 0.053,
    "args": {
      "External id": 539969,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_fused_adamw_", "pid": 4183438, "tid": 4183438,
    "ts": 679622819020.542, "dur": 2606.828,
    "args": {
      "External id": 539970,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "0.00025288793411833431", "0.90000000000000002", "0.94999999999999996", "0.10000000000000001", "1.0000000000000001e-15", "False", "False", "", ""], "Input type": ["TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 9521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_fused_adamw_", "pid": 4183438, "tid": 4183438,
    "ts": 679622821287.626, "dur": 218.428,
    "args": {
      "External id": 539971,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "0.00025288793411833431", "0.90000000000000002", "0.94999999999999996", "0.10000000000000001", "1.0000000000000001e-15", "False", "False", "", ""], "Input type": ["TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 9522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622821807.787, "dur": 16.479,
    "args": {
      "External id": 539972,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622821810.152, "dur": 13.904,
    "args": {
      "External id": 539973,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 4183438,
    "ts": 679622821880.231, "dur": 10.071,
    "args": {
      "External id": 539974,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622821885.024, "dur": 3.300,
    "args": {
      "External id": 539975,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8], [], [], []], "Ev Idx": 9526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622821894.957, "dur": 7.558,
    "args": {
      "External id": 539976,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622821898.280, "dur": 4.019,
    "args": {
      "External id": 539977,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unbind", "pid": 4183438, "tid": 4183438,
    "ts": 679622821921.388, "dur": 13.012,
    "args": {
      "External id": 539978,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[8, 1], []], "Input Dims": [[1, 8], []], "Ev Idx": 9529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622821926.800, "dur": 6.390,
    "args": {
      "External id": 539979,"Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[8, 1], [], []], "Input Dims": [[1, 8], [], []], "Ev Idx": 9530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622821931.959, "dur": 0.499,
    "args": {
      "External id": 539980,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8, 1], [], [], []], "Input Dims": [[1, 8], [], [], []], "Ev Idx": 9531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622821947.155, "dur": 2.113,
    "args": {
      "External id": 539981,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622821947.682, "dur": 1.464,
    "args": {
      "External id": 539982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622821953.995, "dur": 0.587,
    "args": {
      "External id": 539983,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[8], [], [], [], []], "Ev Idx": 9534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 4183438, "tid": 4183438,
    "ts": 679622821960.513, "dur": 0.780,
    "args": {
      "External id": 539984,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "-1"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[8], [], []], "Ev Idx": 9535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622821964.282, "dur": 0.920,
    "args": {
      "External id": 539985,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622821967.177, "dur": 0.260,
    "args": {
      "External id": 539986,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622821986.516, "dur": 36.858,
    "args": {
      "External id": 539987,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622821998.515, "dur": 18.404,
    "args": {
      "External id": 539988,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622821999.900, "dur": 16.849,
    "args": {
      "External id": 539989,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622822004.228, "dur": 2.532,
    "args": {
      "External id": 539990,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622822008.121, "dur": 8.293,
    "args": {
      "External id": 539991,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622822029.324, "dur": 22.891,
    "args": {
      "External id": 539992,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622822032.932, "dur": 1.005,
    "args": {
      "External id": 539993,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided_", "pid": 4183438, "tid": 4183438,
    "ts": 679622822044.046, "dur": 4.430,
    "args": {
      "External id": 539994,"Record function id": 0, "Concrete Inputs": ["", "[1, 1]", "[1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622822057.427, "dur": 4.526,
    "args": {
      "External id": 539995,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[1, 1], [], []], "Ev Idx": 9546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622822060.386, "dur": 0.503,
    "args": {
      "External id": 539996,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622822063.326, "dur": 0.159,
    "args": {
      "External id": 539997,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622822063.908, "dur": 0.218,
    "args": {
      "External id": 539998,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622822079.141, "dur": 7.965,
    "args": {
      "External id": 539999,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622822080.952, "dur": 4.587,
    "args": {
      "External id": 540000,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622822081.494, "dur": 3.905,
    "args": {
      "External id": 540001,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622822082.698, "dur": 0.897,
    "args": {
      "External id": 540002,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622822084.390, "dur": 0.849,
    "args": {
      "External id": 540003,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::any", "pid": 4183438, "tid": 4183438,
    "ts": 679622822092.284, "dur": 31.732,
    "args": {
      "External id": 540004,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622822114.679, "dur": 0.786,
    "args": {
      "External id": 540005,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["bool", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622822121.112, "dur": 1.322,
    "args": {
      "External id": 540006,"Record function id": 0, "Concrete Inputs": ["", "0."], "Input type": ["bool", "Scalar"], "Input Strides": [[0], []], "Input Dims": [[1], []], "Ev Idx": 9557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622822127.214, "dur": 2.685,
    "args": {
      "External id": 540007,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622822128.527, "dur": 1.283,
    "args": {
      "External id": 540008,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622822175.364, "dur": 446.538,
    "args": {
      "External id": 540009,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622822178.271, "dur": 70.950,
    "args": {
      "External id": 540010,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 9561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622822180.660, "dur": 19.488,
    "args": {
      "External id": 540011,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 9562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622822185.298, "dur": 14.276,
    "args": {
      "External id": 540012,"Record function id": 0, "Concrete Inputs": ["[1]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622822201.043, "dur": 47.500,
    "args": {
      "External id": 540013,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[1], [1], []], "Ev Idx": 9564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 4183438, "tid": 4183438,
    "ts": 679622822254.637, "dur": 364.286,
    "args": {
      "External id": 540014,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[1]], [], [], [], []], "Input Dims": [[[1]], [], [], [], []], "Ev Idx": 9565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622822282.618, "dur": 330.406,
    "args": {
      "External id": 540015,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[[1]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[1]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 9566, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622822297.404, "dur": 308.389,
    "args": {
      "External id": 540016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622822708.970, "dur": 1779.972,
    "args": {
      "External id": 540017,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622822712.295, "dur": 1775.967,
    "args": {
      "External id": 540018,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622822780.773, "dur": 31.928,
    "args": {
      "External id": 540019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622822797.657, "dur": 5.987,
    "args": {
      "External id": 540020,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 9571, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622822839.607, "dur": 1636.130,
    "args": {
      "External id": 540021,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 4183438,
    "ts": 679622824537.769, "dur": 9.088,
    "args": {
      "External id": 540022,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622824542.462, "dur": 2.166,
    "args": {
      "External id": 540023,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8], [], [], []], "Ev Idx": 9574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622824550.915, "dur": 5.969,
    "args": {
      "External id": 540024,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622824552.376, "dur": 4.337,
    "args": {
      "External id": 540025,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unbind", "pid": 4183438, "tid": 4183438,
    "ts": 679622824569.053, "dur": 9.288,
    "args": {
      "External id": 540026,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[8, 1], []], "Input Dims": [[1, 8], []], "Ev Idx": 9577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622824572.125, "dur": 5.306,
    "args": {
      "External id": 540027,"Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[8, 1], [], []], "Input Dims": [[1, 8], [], []], "Ev Idx": 9578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622824576.242, "dur": 0.422,
    "args": {
      "External id": 540028,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8, 1], [], [], []], "Input Dims": [[1, 8], [], [], []], "Ev Idx": 9579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622824589.657, "dur": 4.214,
    "args": {
      "External id": 540029,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622824591.059, "dur": 2.630,
    "args": {
      "External id": 540030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622824597.324, "dur": 0.747,
    "args": {
      "External id": 540031,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[8], [], [], [], []], "Ev Idx": 9582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 4183438, "tid": 4183438,
    "ts": 679622824601.542, "dur": 0.310,
    "args": {
      "External id": 540032,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "-1"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[8], [], []], "Ev Idx": 9583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622824603.719, "dur": 0.561,
    "args": {
      "External id": 540033,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622824605.386, "dur": 0.171,
    "args": {
      "External id": 540034,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622824619.350, "dur": 28.241,
    "args": {
      "External id": 540035,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622824627.761, "dur": 15.330,
    "args": {
      "External id": 540036,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622824629.127, "dur": 13.774,
    "args": {
      "External id": 540037,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622824633.050, "dur": 2.480,
    "args": {
      "External id": 540038,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622824636.696, "dur": 5.880,
    "args": {
      "External id": 540039,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622824691.625, "dur": 16.562,
    "args": {
      "External id": 540040,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622824695.504, "dur": 1.576,
    "args": {
      "External id": 540041,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided_", "pid": 4183438, "tid": 4183438,
    "ts": 679622824703.856, "dur": 2.917,
    "args": {
      "External id": 540042,"Record function id": 0, "Concrete Inputs": ["", "[1, 1]", "[1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622824715.811, "dur": 5.336,
    "args": {
      "External id": 540043,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[1, 1], [], []], "Ev Idx": 9594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622824719.120, "dur": 0.634,
    "args": {
      "External id": 540044,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622824722.811, "dur": 0.280,
    "args": {
      "External id": 540045,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622824723.858, "dur": 0.249,
    "args": {
      "External id": 540046,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622824733.017, "dur": 11.113,
    "args": {
      "External id": 540047,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622824735.490, "dur": 6.805,
    "args": {
      "External id": 540048,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622824736.461, "dur": 5.639,
    "args": {
      "External id": 540049,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622824737.953, "dur": 1.836,
    "args": {
      "External id": 540050,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622824740.488, "dur": 1.418,
    "args": {
      "External id": 540051,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::any", "pid": 4183438, "tid": 4183438,
    "ts": 679622824746.782, "dur": 9.051,
    "args": {
      "External id": 540052,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622824750.781, "dur": 0.413,
    "args": {
      "External id": 540053,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["bool", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622824754.404, "dur": 0.633,
    "args": {
      "External id": 540054,"Record function id": 0, "Concrete Inputs": ["", "0."], "Input type": ["bool", "Scalar"], "Input Strides": [[0], []], "Input Dims": [[1], []], "Ev Idx": 9605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622824757.652, "dur": 1.473,
    "args": {
      "External id": 540055,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622824758.353, "dur": 0.694,
    "args": {
      "External id": 540056,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622824801.739, "dur": 429.205,
    "args": {
      "External id": 540057,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622824804.292, "dur": 50.593,
    "args": {
      "External id": 540058,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 9609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622824806.598, "dur": 12.030,
    "args": {
      "External id": 540059,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 9610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622824808.803, "dur": 9.357,
    "args": {
      "External id": 540060,"Record function id": 0, "Concrete Inputs": ["[1]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622824819.492, "dur": 34.834,
    "args": {
      "External id": 540061,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[1], [1], []], "Ev Idx": 9612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 4183438, "tid": 4183438,
    "ts": 679622824859.857, "dur": 368.313,
    "args": {
      "External id": 540062,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[1]], [], [], [], []], "Input Dims": [[[1]], [], [], [], []], "Ev Idx": 9613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622824884.222, "dur": 337.493,
    "args": {
      "External id": 540063,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[[1]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[1]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 9614, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622824900.383, "dur": 315.918,
    "args": {
      "External id": 540064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622825276.276, "dur": 1060.112,
    "args": {
      "External id": 540065,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622825278.591, "dur": 1057.217,
    "args": {
      "External id": 540066,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622825340.065, "dur": 27.001,
    "args": {
      "External id": 540067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622825355.169, "dur": 4.165,
    "args": {
      "External id": 540068,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 9619, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622825391.465, "dur": 932.527,
    "args": {
      "External id": 540069,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 4183438,
    "ts": 679622826384.141, "dur": 8.599,
    "args": {
      "External id": 540070,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622826388.447, "dur": 2.338,
    "args": {
      "External id": 540071,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8], [], [], []], "Ev Idx": 9622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622826396.666, "dur": 6.526,
    "args": {
      "External id": 540072,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622826398.414, "dur": 4.577,
    "args": {
      "External id": 540073,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unbind", "pid": 4183438, "tid": 4183438,
    "ts": 679622826414.189, "dur": 9.379,
    "args": {
      "External id": 540074,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[8, 1], []], "Input Dims": [[1, 8], []], "Ev Idx": 9625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622826417.266, "dur": 5.362,
    "args": {
      "External id": 540075,"Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[8, 1], [], []], "Input Dims": [[1, 8], [], []], "Ev Idx": 9626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622826421.323, "dur": 0.484,
    "args": {
      "External id": 540076,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8, 1], [], [], []], "Input Dims": [[1, 8], [], [], []], "Ev Idx": 9627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622826434.572, "dur": 4.478,
    "args": {
      "External id": 540077,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622826436.065, "dur": 2.717,
    "args": {
      "External id": 540078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622826442.433, "dur": 0.636,
    "args": {
      "External id": 540079,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[8], [], [], [], []], "Ev Idx": 9630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 4183438, "tid": 4183438,
    "ts": 679622826446.032, "dur": 0.430,
    "args": {
      "External id": 540080,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "-1"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[8], [], []], "Ev Idx": 9631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622826448.585, "dur": 0.436,
    "args": {
      "External id": 540081,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622826449.804, "dur": 0.121,
    "args": {
      "External id": 540082,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622826461.638, "dur": 31.760,
    "args": {
      "External id": 540083,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622826469.857, "dur": 19.228,
    "args": {
      "External id": 540084,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622826471.153, "dur": 17.758,
    "args": {
      "External id": 540085,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622826478.625, "dur": 2.949,
    "args": {
      "External id": 540086,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622826482.890, "dur": 5.697,
    "args": {
      "External id": 540087,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622826496.365, "dur": 15.081,
    "args": {
      "External id": 540088,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622826498.577, "dur": 0.870,
    "args": {
      "External id": 540089,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided_", "pid": 4183438, "tid": 4183438,
    "ts": 679622826507.709, "dur": 2.453,
    "args": {
      "External id": 540090,"Record function id": 0, "Concrete Inputs": ["", "[1, 1]", "[1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622826515.975, "dur": 4.312,
    "args": {
      "External id": 540091,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[1, 1], [], []], "Ev Idx": 9642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622826518.637, "dur": 0.437,
    "args": {
      "External id": 540092,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622826521.531, "dur": 0.109,
    "args": {
      "External id": 540093,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622826522.111, "dur": 0.080,
    "args": {
      "External id": 540094,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622826529.389, "dur": 8.214,
    "args": {
      "External id": 540095,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622826531.338, "dur": 4.893,
    "args": {
      "External id": 540096,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622826532.058, "dur": 4.021,
    "args": {
      "External id": 540097,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622826533.336, "dur": 1.186,
    "args": {
      "External id": 540098,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622826535.084, "dur": 0.847,
    "args": {
      "External id": 540099,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::any", "pid": 4183438, "tid": 4183438,
    "ts": 679622826539.990, "dur": 10.725,
    "args": {
      "External id": 540100,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622826543.565, "dur": 2.485,
    "args": {
      "External id": 540101,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["bool", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622826549.307, "dur": 0.607,
    "args": {
      "External id": 540102,"Record function id": 0, "Concrete Inputs": ["", "0."], "Input type": ["bool", "Scalar"], "Input Strides": [[0], []], "Input Dims": [[1], []], "Ev Idx": 9653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622826552.827, "dur": 1.462,
    "args": {
      "External id": 540103,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622826553.543, "dur": 0.674,
    "args": {
      "External id": 540104,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622826587.377, "dur": 449.600,
    "args": {
      "External id": 540105,"Sequence number": 6319638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622826593.365, "dur": 50.949,
    "args": {
      "External id": 540106,"Sequence number": 6319639, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622826596.843, "dur": 12.505,
    "args": {
      "External id": 540107,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622826598.883, "dur": 9.905,
    "args": {
      "External id": 540108,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622826610.346, "dur": 31.503,
    "args": {
      "External id": 540109,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 4183438, "tid": 4183438,
    "ts": 679622826649.195, "dur": 383.914,
    "args": {
      "External id": 540110,"Sequence number": 6319640, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[]], [], [], [], []], "Input Dims": [[[]], [], [], [], []], "Ev Idx": 9661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622826706.926, "dur": 315.099,
    "args": {
      "External id": 540111,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[[]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 9662, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622826724.711, "dur": 292.588,
    "args": {
      "External id": 540112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622827080.555, "dur": 997.989,
    "args": {
      "External id": 540113,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622827082.533, "dur": 995.292,
    "args": {
      "External id": 540114,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622827160.634, "dur": 28.575,
    "args": {
      "External id": 540115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622827175.350, "dur": 4.998,
    "args": {
      "External id": 540116,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 9667, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622827213.460, "dur": 852.090,
    "args": {
      "External id": 540117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 4183438,
    "ts": 679622828141.722, "dur": 8.501,
    "args": {
      "External id": 540118,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622828146.082, "dur": 2.168,
    "args": {
      "External id": 540119,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8], [], [], []], "Ev Idx": 9670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622828154.110, "dur": 6.074,
    "args": {
      "External id": 540120,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622828155.586, "dur": 4.413,
    "args": {
      "External id": 540121,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unbind", "pid": 4183438, "tid": 4183438,
    "ts": 679622828170.875, "dur": 9.609,
    "args": {
      "External id": 540122,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[8, 1], []], "Input Dims": [[1, 8], []], "Ev Idx": 9673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622828173.819, "dur": 5.778,
    "args": {
      "External id": 540123,"Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[8, 1], [], []], "Input Dims": [[1, 8], [], []], "Ev Idx": 9674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622828178.331, "dur": 0.465,
    "args": {
      "External id": 540124,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8, 1], [], [], []], "Input Dims": [[1, 8], [], [], []], "Ev Idx": 9675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622828191.450, "dur": 4.450,
    "args": {
      "External id": 540125,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622828192.850, "dur": 2.827,
    "args": {
      "External id": 540126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622828199.600, "dur": 0.516,
    "args": {
      "External id": 540127,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[8], [], [], [], []], "Ev Idx": 9678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 4183438, "tid": 4183438,
    "ts": 679622828202.798, "dur": 0.529,
    "args": {
      "External id": 540128,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "-1"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[8], [], []], "Ev Idx": 9679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622828205.107, "dur": 0.398,
    "args": {
      "External id": 540129,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622828206.281, "dur": 0.274,
    "args": {
      "External id": 540130,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622828218.148, "dur": 27.485,
    "args": {
      "External id": 540131,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622828226.296, "dur": 15.106,
    "args": {
      "External id": 540132,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622828227.592, "dur": 13.619,
    "args": {
      "External id": 540133,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622828231.514, "dur": 2.394,
    "args": {
      "External id": 540134,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622828235.133, "dur": 5.740,
    "args": {
      "External id": 540135,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622828248.635, "dur": 12.396,
    "args": {
      "External id": 540136,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622828250.721, "dur": 0.797,
    "args": {
      "External id": 540137,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided_", "pid": 4183438, "tid": 4183438,
    "ts": 679622828257.246, "dur": 2.515,
    "args": {
      "External id": 540138,"Record function id": 0, "Concrete Inputs": ["", "[1, 1]", "[1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622828265.869, "dur": 4.341,
    "args": {
      "External id": 540139,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[1, 1], [], []], "Ev Idx": 9690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622828268.742, "dur": 0.439,
    "args": {
      "External id": 540140,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622828271.510, "dur": 0.224,
    "args": {
      "External id": 540141,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622828272.519, "dur": 0.202,
    "args": {
      "External id": 540142,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622828279.644, "dur": 8.709,
    "args": {
      "External id": 540143,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622828281.333, "dur": 5.191,
    "args": {
      "External id": 540144,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622828282.151, "dur": 4.189,
    "args": {
      "External id": 540145,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622828283.363, "dur": 1.231,
    "args": {
      "External id": 540146,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622828285.361, "dur": 0.813,
    "args": {
      "External id": 540147,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::any", "pid": 4183438, "tid": 4183438,
    "ts": 679622828290.831, "dur": 8.784,
    "args": {
      "External id": 540148,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622828294.637, "dur": 0.516,
    "args": {
      "External id": 540149,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["bool", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622828298.151, "dur": 0.628,
    "args": {
      "External id": 540150,"Record function id": 0, "Concrete Inputs": ["", "0."], "Input type": ["bool", "Scalar"], "Input Strides": [[0], []], "Input Dims": [[1], []], "Ev Idx": 9701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622828301.353, "dur": 1.795,
    "args": {
      "External id": 540151,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622828302.133, "dur": 0.929,
    "args": {
      "External id": 540152,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622828337.865, "dur": 461.041,
    "args": {
      "External id": 540153,"Sequence number": 6319641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622828342.695, "dur": 48.036,
    "args": {
      "External id": 540154,"Sequence number": 6319642, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 9705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622828346.061, "dur": 10.201,
    "args": {
      "External id": 540155,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622828347.967, "dur": 7.867,
    "args": {
      "External id": 540156,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622828357.133, "dur": 31.479,
    "args": {
      "External id": 540157,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 4183438, "tid": 4183438,
    "ts": 679622828398.355, "dur": 397.252,
    "args": {
      "External id": 540158,"Sequence number": 6319643, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[]], [], [], [], []], "Input Dims": [[[]], [], [], [], []], "Ev Idx": 9709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622828422.290, "dur": 362.349,
    "args": {
      "External id": 540159,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[[]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 9710, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622828437.650, "dur": 340.857,
    "args": {
      "External id": 540160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622828865.242, "dur": 1027.071,
    "args": {
      "External id": 540161,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622828867.498, "dur": 1023.905,
    "args": {
      "External id": 540162,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622828927.548, "dur": 28.367,
    "args": {
      "External id": 540163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622828943.981, "dur": 4.227,
    "args": {
      "External id": 540164,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 9715, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622828979.810, "dur": 898.828,
    "args": {
      "External id": 540165,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 4183438,
    "ts": 679622829942.630, "dur": 8.328,
    "args": {
      "External id": 540166,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622829946.976, "dur": 2.103,
    "args": {
      "External id": 540167,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8], [], [], []], "Ev Idx": 9718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622829954.716, "dur": 5.594,
    "args": {
      "External id": 540168,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622829956.142, "dur": 3.983,
    "args": {
      "External id": 540169,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unbind", "pid": 4183438, "tid": 4183438,
    "ts": 679622829970.409, "dur": 8.701,
    "args": {
      "External id": 540170,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[8, 1], []], "Input Dims": [[1, 8], []], "Ev Idx": 9721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622829972.859, "dur": 5.380,
    "args": {
      "External id": 540171,"Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[8, 1], [], []], "Input Dims": [[1, 8], [], []], "Ev Idx": 9722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622829977.072, "dur": 0.414,
    "args": {
      "External id": 540172,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8, 1], [], [], []], "Input Dims": [[1, 8], [], [], []], "Ev Idx": 9723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622829990.613, "dur": 4.099,
    "args": {
      "External id": 540173,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622829991.866, "dur": 2.688,
    "args": {
      "External id": 540174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622829998.407, "dur": 0.473,
    "args": {
      "External id": 540175,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[8], [], [], [], []], "Ev Idx": 9726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 4183438, "tid": 4183438,
    "ts": 679622830001.757, "dur": 0.397,
    "args": {
      "External id": 540176,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "-1"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[8], [], []], "Ev Idx": 9727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622830004.010, "dur": 0.533,
    "args": {
      "External id": 540177,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622830005.571, "dur": 0.256,
    "args": {
      "External id": 540178,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622830018.158, "dur": 27.087,
    "args": {
      "External id": 540179,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622830026.380, "dur": 14.700,
    "args": {
      "External id": 540180,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622830027.608, "dur": 13.302,
    "args": {
      "External id": 540181,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622830031.504, "dur": 2.347,
    "args": {
      "External id": 540182,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622830035.026, "dur": 5.607,
    "args": {
      "External id": 540183,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622830047.882, "dur": 12.768,
    "args": {
      "External id": 540184,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622830049.871, "dur": 0.893,
    "args": {
      "External id": 540185,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided_", "pid": 4183438, "tid": 4183438,
    "ts": 679622830056.830, "dur": 2.600,
    "args": {
      "External id": 540186,"Record function id": 0, "Concrete Inputs": ["", "[1, 1]", "[1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622830065.182, "dur": 4.303,
    "args": {
      "External id": 540187,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[1, 1], [], []], "Ev Idx": 9738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622830067.826, "dur": 0.559,
    "args": {
      "External id": 540188,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622830070.614, "dur": 0.077,
    "args": {
      "External id": 540189,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622830071.107, "dur": 0.073,
    "args": {
      "External id": 540190,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622830078.056, "dur": 7.811,
    "args": {
      "External id": 540191,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622830079.771, "dur": 4.572,
    "args": {
      "External id": 540192,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622830080.197, "dur": 3.966,
    "args": {
      "External id": 540193,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622830081.467, "dur": 1.175,
    "args": {
      "External id": 540194,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622830083.212, "dur": 0.798,
    "args": {
      "External id": 540195,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::any", "pid": 4183438, "tid": 4183438,
    "ts": 679622830088.258, "dur": 8.450,
    "args": {
      "External id": 540196,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622830091.945, "dur": 0.373,
    "args": {
      "External id": 540197,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["bool", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622830095.327, "dur": 0.576,
    "args": {
      "External id": 540198,"Record function id": 0, "Concrete Inputs": ["", "0."], "Input type": ["bool", "Scalar"], "Input Strides": [[0], []], "Input Dims": [[1], []], "Ev Idx": 9749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622830098.660, "dur": 16.340,
    "args": {
      "External id": 540199,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622830099.370, "dur": 15.160,
    "args": {
      "External id": 540200,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622830149.638, "dur": 425.314,
    "args": {
      "External id": 540201,"Sequence number": 6319644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622830155.151, "dur": 50.029,
    "args": {
      "External id": 540202,"Sequence number": 6319645, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 9753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622830158.419, "dur": 11.316,
    "args": {
      "External id": 540203,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 9754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622830160.584, "dur": 8.743,
    "args": {
      "External id": 540204,"Record function id": 0, "Concrete Inputs": ["[1]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622830170.462, "dur": 32.224,
    "args": {
      "External id": 540205,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[1], [1], []], "Ev Idx": 9756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 4183438, "tid": 4183438,
    "ts": 679622830210.590, "dur": 361.132,
    "args": {
      "External id": 540206,"Sequence number": 6319646, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[1]], [], [], [], []], "Input Dims": [[[1]], [], [], [], []], "Ev Idx": 9757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622830236.509, "dur": 325.696,
    "args": {
      "External id": 540207,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[[1]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[1]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 9758, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622830250.794, "dur": 306.197,
    "args": {
      "External id": 540208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622830617.209, "dur": 989.648,
    "args": {
      "External id": 540209,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622830619.081, "dur": 986.998,
    "args": {
      "External id": 540210,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622830712.744, "dur": 29.546,
    "args": {
      "External id": 540211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622830728.202, "dur": 5.125,
    "args": {
      "External id": 540212,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 9763, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622830766.640, "dur": 827.428,
    "args": {
      "External id": 540213,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 4183438, "tid": 4183438,
    "ts": 679622831653.405, "dur": 35.506,
    "args": {
      "External id": 540214,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622831684.619, "dur": 2.158,
    "args": {
      "External id": 540215,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8], [], [], []], "Ev Idx": 9766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 4183438, "tid": 4183438,
    "ts": 679622831693.633, "dur": 6.157,
    "args": {
      "External id": 540216,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 4183438, "tid": 4183438,
    "ts": 679622831695.444, "dur": 4.145,
    "args": {
      "External id": 540217,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8]"], "Input type": ["int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unbind", "pid": 4183438, "tid": 4183438,
    "ts": 679622831710.566, "dur": 8.684,
    "args": {
      "External id": 540218,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[8, 1], []], "Input Dims": [[1, 8], []], "Ev Idx": 9769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622831713.205, "dur": 5.150,
    "args": {
      "External id": 540219,"Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[8, 1], [], []], "Input Dims": [[1, 8], [], []], "Ev Idx": 9770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622831717.121, "dur": 0.529,
    "args": {
      "External id": 540220,"Record function id": 0, "Concrete Inputs": ["", "[8]", "[1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8, 1], [], [], []], "Input Dims": [[1, 8], [], [], []], "Ev Idx": 9771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622831730.484, "dur": 4.240,
    "args": {
      "External id": 540221,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 4183438, "tid": 4183438,
    "ts": 679622831731.773, "dur": 2.763,
    "args": {
      "External id": 540222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622831737.657, "dur": 0.482,
    "args": {
      "External id": 540223,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[8], [], [], [], []], "Ev Idx": 9774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 4183438, "tid": 4183438,
    "ts": 679622831740.998, "dur": 0.440,
    "args": {
      "External id": 540224,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "-1"], "Input type": ["int", "Scalar", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[8], [], []], "Ev Idx": 9775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622831743.416, "dur": 0.462,
    "args": {
      "External id": 540225,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622831744.738, "dur": 0.203,
    "args": {
      "External id": 540226,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622831756.352, "dur": 27.128,
    "args": {
      "External id": 540227,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622831763.930, "dur": 15.098,
    "args": {
      "External id": 540228,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622831765.133, "dur": 13.692,
    "args": {
      "External id": 540229,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622831769.193, "dur": 2.549,
    "args": {
      "External id": 540230,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622831772.869, "dur": 5.661,
    "args": {
      "External id": 540231,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 4183438, "tid": 4183438,
    "ts": 679622831786.075, "dur": 12.133,
    "args": {
      "External id": 540232,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622831787.975, "dur": 0.860,
    "args": {
      "External id": 540233,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided_", "pid": 4183438, "tid": 4183438,
    "ts": 679622831794.495, "dur": 2.532,
    "args": {
      "External id": 540234,"Record function id": 0, "Concrete Inputs": ["", "[1, 1]", "[1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 4183438, "tid": 4183438,
    "ts": 679622831802.501, "dur": 4.400,
    "args": {
      "External id": 540235,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[1, 1], [], []], "Ev Idx": 9786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622831805.446, "dur": 0.304,
    "args": {
      "External id": 540236,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[1, 1], [], [], []], "Ev Idx": 9787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 4183438, "tid": 4183438,
    "ts": 679622831808.059, "dur": 0.078,
    "args": {
      "External id": 540237,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 4183438, "tid": 4183438,
    "ts": 679622831808.564, "dur": 0.197,
    "args": {
      "External id": 540238,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 4183438, "tid": 4183438,
    "ts": 679622831815.674, "dur": 7.934,
    "args": {
      "External id": 540239,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8], []], "Ev Idx": 9790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 4183438, "tid": 4183438,
    "ts": 679622831817.631, "dur": 4.501,
    "args": {
      "External id": 540240,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 4183438, "tid": 4183438,
    "ts": 679622831818.270, "dur": 3.693,
    "args": {
      "External id": 540241,"Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 9792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622831819.534, "dur": 0.955,
    "args": {
      "External id": 540242,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622831820.983, "dur": 0.808,
    "args": {
      "External id": 540243,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 9794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::any", "pid": 4183438, "tid": 4183438,
    "ts": 679622831826.028, "dur": 8.549,
    "args": {
      "External id": 540244,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[1]], "Input Dims": [[8]], "Ev Idx": 9795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 4183438, "tid": 4183438,
    "ts": 679622831829.609, "dur": 0.435,
    "args": {
      "External id": 540245,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["bool", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 4183438, "tid": 4183438,
    "ts": 679622831833.074, "dur": 0.676,
    "args": {
      "External id": 540246,"Record function id": 0, "Concrete Inputs": ["", "0."], "Input type": ["bool", "Scalar"], "Input Strides": [[0], []], "Input Dims": [[1], []], "Ev Idx": 9797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622831836.516, "dur": 1.840,
    "args": {
      "External id": 540247,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622831837.402, "dur": 0.869,
    "args": {
      "External id": 540248,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622831872.713, "dur": 429.783,
    "args": {
      "External id": 540249,"Sequence number": 6319647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 4183438, "tid": 4183438,
    "ts": 679622831877.607, "dur": 48.099,
    "args": {
      "External id": 540250,"Sequence number": 6319648, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 9801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 4183438, "tid": 4183438,
    "ts": 679622831880.976, "dur": 11.061,
    "args": {
      "External id": 540251,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 9802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 4183438, "tid": 4183438,
    "ts": 679622831883.353, "dur": 8.250,
    "args": {
      "External id": 540252,"Record function id": 0, "Concrete Inputs": ["[1]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 4183438, "tid": 4183438,
    "ts": 679622831892.862, "dur": 30.636,
    "args": {
      "External id": 540253,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[1], [1], []], "Ev Idx": 9804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 4183438, "tid": 4183438,
    "ts": 679622831931.382, "dur": 367.804,
    "args": {
      "External id": 540254,"Sequence number": 6319649, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[1]], [], [], [], []], "Input Dims": [[[1]], [], [], [], []], "Ev Idx": 9805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622831954.871, "dur": 334.143,
    "args": {
      "External id": 540255,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 2, "Input Strides": [[[1]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "2", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[1]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 9806, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 4183438, "tid": 4183438,
    "ts": 679622831971.068, "dur": 312.127,
    "args": {
      "External id": 540256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622832348.361, "dur": 988.845,
    "args": {
      "External id": 540257,"Sequence number": 6319650, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622832350.364, "dur": 985.925,
    "args": {
      "External id": 540258,"Sequence number": 6319650, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 4183438, "tid": 4183438,
    "ts": 679622832411.177, "dur": 27.460,
    "args": {
      "External id": 540259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 4183438, "tid": 4183438,
    "ts": 679622832426.657, "dur": 4.466,
    "args": {
      "External id": 540260,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "2", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 9811, "In msg nelems": 0, "Rank": 2, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622832464.345, "dur": 859.267,
    "args": {
      "External id": 540261,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 9812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622833381.055, "dur": 35.165,
    "args": {
      "External id": 540262,"Sequence number": 6319650, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622833382.186, "dur": 33.756,
    "args": {
      "External id": 540263,"Sequence number": 6319650, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 4183438, "tid": 4183438,
    "ts": 679622833868.655, "dur": 43.788,
    "args": {
      "External id": 540264,"Sequence number": 6319650, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 4183438, "tid": 4183438,
    "ts": 679622833869.564, "dur": 42.148,
    "args": {
      "External id": 540265,"Sequence number": 6319650, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9816
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 0,
    "args": {
      "name": "python3.12"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 0,
    "args": {
      "sort_index": 4183438
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 31367,
    "args": {
      "name": "thread 31367 (pt_autograd_2)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 31367,
    "args": {
      "sort_index": 31367
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 31367,
    "args": {
      "name": "thread 31367 (python3.12)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 31367,
    "args": {
      "sort_index": 31367
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 4183438,
    "args": {
      "name": "thread 4183438 (python3.12)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 679622021896.159, "pid": 4183438, "tid": 4183438,
    "args": {
      "sort_index": 4183438
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 679622021828.725, "dur": 812342.599,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 679622021828.725,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 679622021828.725
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 679622882524.699
  }
  ],
  "traceName": "exp/mtp.340M.batch16.seqlen4096.context4096.warmup1000.update1.steps100000.lr3e-4.cosine/profile_trace/iteration_28160/rank2_trace.json",
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1751410836000000000
}