Add files using upload-large-folder tool

l2-13b-ga/checkpoint-2800/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

l2-13b-ga/checkpoint-2800/tokenizer_config.json

@@ -0,0 +1,42 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": true,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": null,
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

l2-13b-ga/checkpoint-3100/config.json

@@ -0,0 +1,30 @@
+{
+  "_name_or_path": "meta-llama/Llama-2-13b-hf",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 13824,
+  "max_position_embeddings": 4096,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 40,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.46.3",
+  "use_cache": true,
+  "vocab_size": 35483
+}

l2-13b-ga/checkpoint-3100/generation_config.json

@@ -0,0 +1,10 @@
+{
+  "bos_token_id": 1,
+  "do_sample": true,
+  "eos_token_id": 2,
+  "max_length": 4096,
+  "pad_token_id": 0,
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.46.3"
+}

l2-13b-ga/checkpoint-3100/latest

@@ -0,0 +1 @@
+global_step3100

l2-13b-ga/checkpoint-3100/model.safetensors.index.json

@@ -0,0 +1,370 @@
+{
+  "metadata": {
+    "total_size": 26103060480
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00006-of-00006.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.22.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.30.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.30.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.30.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.37.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.38.input_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.mlp.gate_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.mlp.up_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.38.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.38.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.38.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.39.input_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.mlp.gate_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.mlp.up_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.self_attn.k_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.self_attn.q_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.self_attn.v_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
+    "model.norm.weight": "model-00006-of-00006.safetensors"
+  }
+}

l2-13b-ga/checkpoint-3100/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

l2-13b-ga/checkpoint-3100/tokenizer_config.json

@@ -0,0 +1,42 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": true,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": null,
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

l2-13b-ga/checkpoint-3100/trainer_state.json

@@ -0,0 +1,2210 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.8954004202661685,
+  "eval_steps": 500,
+  "global_step": 3100,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0009339248190520663,
+      "grad_norm": 6.638877692627699,
+      "learning_rate": 9.345794392523364e-07,
+      "loss": 9.2917,
+      "step": 1
+    },
+    {
+      "epoch": 0.009339248190520663,
+      "grad_norm": 1.1560921335705272,
+      "learning_rate": 9.345794392523365e-06,
+      "loss": 9.0876,
+      "step": 10
+    },
+    {
+      "epoch": 0.018678496381041326,
+      "grad_norm": 0.8415132296956432,
+      "learning_rate": 1.869158878504673e-05,
+      "loss": 8.2164,
+      "step": 20
+    },
+    {
+      "epoch": 0.02801774457156199,
+      "grad_norm": 0.45381630992958155,
+      "learning_rate": 2.8037383177570094e-05,
+      "loss": 7.5184,
+      "step": 30
+    },
+    {
+      "epoch": 0.03735699276208265,
+      "grad_norm": 0.8400636107958425,
+      "learning_rate": 3.738317757009346e-05,
+      "loss": 6.6507,
+      "step": 40
+    },
+    {
+      "epoch": 0.046696240952603316,
+      "grad_norm": 0.557696240829066,
+      "learning_rate": 4.672897196261683e-05,
+      "loss": 5.8909,
+      "step": 50
+    },
+    {
+      "epoch": 0.05603548914312398,
+      "grad_norm": 0.3971996057467842,
+      "learning_rate": 5.607476635514019e-05,
+      "loss": 5.4127,
+      "step": 60
+    },
+    {
+      "epoch": 0.06537473733364464,
+      "grad_norm": 0.2932710540265688,
+      "learning_rate": 6.542056074766355e-05,
+      "loss": 5.0106,
+      "step": 70
+    },
+    {
+      "epoch": 0.0747139855241653,
+      "grad_norm": 0.3682690443551033,
+      "learning_rate": 7.476635514018692e-05,
+      "loss": 4.6042,
+      "step": 80
+    },
+    {
+      "epoch": 0.08405323371468597,
+      "grad_norm": 0.3132971920011515,
+      "learning_rate": 8.411214953271028e-05,
+      "loss": 4.2031,
+      "step": 90
+    },
+    {
+      "epoch": 0.09339248190520663,
+      "grad_norm": 0.6731868159213446,
+      "learning_rate": 9.345794392523365e-05,
+      "loss": 3.9423,
+      "step": 100
+    },
+    {
+      "epoch": 0.1027317300957273,
+      "grad_norm": 0.27848867836763197,
+      "learning_rate": 0.000102803738317757,
+      "loss": 3.7157,
+      "step": 110
+    },
+    {
+      "epoch": 0.11207097828624796,
+      "grad_norm": 0.24642109032991807,
+      "learning_rate": 0.00011214953271028037,
+      "loss": 3.4516,
+      "step": 120
+    },
+    {
+      "epoch": 0.12141022647676862,
+      "grad_norm": 0.25717384664029797,
+      "learning_rate": 0.00012149532710280373,
+      "loss": 3.2167,
+      "step": 130
+    },
+    {
+      "epoch": 0.13074947466728928,
+      "grad_norm": 0.20912922668565637,
+      "learning_rate": 0.0001308411214953271,
+      "loss": 3.0237,
+      "step": 140
+    },
+    {
+      "epoch": 0.14008872285780993,
+      "grad_norm": 0.15805888388706113,
+      "learning_rate": 0.00014018691588785047,
+      "loss": 2.8529,
+      "step": 150
+    },
+    {
+      "epoch": 0.1494279710483306,
+      "grad_norm": 0.23370349497479534,
+      "learning_rate": 0.00014953271028037384,
+      "loss": 2.7078,
+      "step": 160
+    },
+    {
+      "epoch": 0.15876721923885126,
+      "grad_norm": 0.1802138633012483,
+      "learning_rate": 0.0001588785046728972,
+      "loss": 2.6115,
+      "step": 170
+    },
+    {
+      "epoch": 0.16810646742937194,
+      "grad_norm": 0.13354347610039718,
+      "learning_rate": 0.00016822429906542056,
+      "loss": 2.5309,
+      "step": 180
+    },
+    {
+      "epoch": 0.17744571561989259,
+      "grad_norm": 0.09414865188086892,
+      "learning_rate": 0.00017757009345794393,
+      "loss": 2.4452,
+      "step": 190
+    },
+    {
+      "epoch": 0.18678496381041326,
+      "grad_norm": 0.08333601554768896,
+      "learning_rate": 0.0001869158878504673,
+      "loss": 2.3832,
+      "step": 200
+    },
+    {
+      "epoch": 0.1961242120009339,
+      "grad_norm": 0.15926414699806835,
+      "learning_rate": 0.00019626168224299065,
+      "loss": 2.3492,
+      "step": 210
+    },
+    {
+      "epoch": 0.2054634601914546,
+      "grad_norm": 0.09492820761057012,
+      "learning_rate": 0.0001999989254250208,
+      "loss": 2.323,
+      "step": 220
+    },
+    {
+      "epoch": 0.21480270838197524,
+      "grad_norm": 0.0801349259356147,
+      "learning_rate": 0.00019999235866155886,
+      "loss": 2.2731,
+      "step": 230
+    },
+    {
+      "epoch": 0.22414195657249592,
+      "grad_norm": 0.12210960524693895,
+      "learning_rate": 0.00019997982251228469,
+      "loss": 2.2433,
+      "step": 240
+    },
+    {
+      "epoch": 0.23348120476301656,
+      "grad_norm": 3.14289498732125,
+      "learning_rate": 0.00019996131772558666,
+      "loss": 3.2769,
+      "step": 250
+    },
+    {
+      "epoch": 0.24282045295353724,
+      "grad_norm": 1.632940983166179,
+      "learning_rate": 0.00019993684540617132,
+      "loss": 4.9343,
+      "step": 260
+    },
+    {
+      "epoch": 0.2521597011440579,
+      "grad_norm": 3.4831252230225416,
+      "learning_rate": 0.00019990640701499736,
+      "loss": 4.2768,
+      "step": 270
+    },
+    {
+      "epoch": 0.26149894933457857,
+      "grad_norm": 1.6069045920523788,
+      "learning_rate": 0.00019987000436918874,
+      "loss": 5.9581,
+      "step": 280
+    },
+    {
+      "epoch": 0.27083819752509924,
+      "grad_norm": 0.2220907936615993,
+      "learning_rate": 0.00019982763964192585,
+      "loss": 3.8228,
+      "step": 290
+    },
+    {
+      "epoch": 0.28017744571561987,
+      "grad_norm": 0.24737284913291765,
+      "learning_rate": 0.00019977931536231596,
+      "loss": 3.1413,
+      "step": 300
+    },
+    {
+      "epoch": 0.28951669390614054,
+      "grad_norm": 4.010404518241152,
+      "learning_rate": 0.00019972503441524224,
+      "loss": 2.8432,
+      "step": 310
+    },
+    {
+      "epoch": 0.2988559420966612,
+      "grad_norm": 0.1515583580811596,
+      "learning_rate": 0.00019966480004119142,
+      "loss": 2.7859,
+      "step": 320
+    },
+    {
+      "epoch": 0.3081951902871819,
+      "grad_norm": 0.11259395750650594,
+      "learning_rate": 0.00019959861583606045,
+      "loss": 2.5821,
+      "step": 330
+    },
+    {
+      "epoch": 0.3175344384777025,
+      "grad_norm": 0.22514797814956813,
+      "learning_rate": 0.00019952648575094183,
+      "loss": 2.4517,
+      "step": 340
+    },
+    {
+      "epoch": 0.3268736866682232,
+      "grad_norm": 0.08040136172033542,
+      "learning_rate": 0.00019944841409188767,
+      "loss": 2.3794,
+      "step": 350
+    },
+    {
+      "epoch": 0.3362129348587439,
+      "grad_norm": 0.054758073593565354,
+      "learning_rate": 0.00019936440551965263,
+      "loss": 2.3232,
+      "step": 360
+    },
+    {
+      "epoch": 0.34555218304926455,
+      "grad_norm": 0.06742998909645591,
+      "learning_rate": 0.00019927446504941577,
+      "loss": 2.2776,
+      "step": 370
+    },
+    {
+      "epoch": 0.35489143123978517,
+      "grad_norm": 0.048780907584876736,
+      "learning_rate": 0.00019917859805048096,
+      "loss": 2.2376,
+      "step": 380
+    },
+    {
+      "epoch": 0.36423067943030585,
+      "grad_norm": 0.0475325963052214,
+      "learning_rate": 0.00019907681024595663,
+      "loss": 2.2191,
+      "step": 390
+    },
+    {
+      "epoch": 0.3735699276208265,
+      "grad_norm": 0.054089563211590065,
+      "learning_rate": 0.00019896910771241387,
+      "loss": 2.1961,
+      "step": 400
+    },
+    {
+      "epoch": 0.3829091758113472,
+      "grad_norm": 0.21798406131864823,
+      "learning_rate": 0.00019885549687952372,
+      "loss": 2.2078,
+      "step": 410
+    },
+    {
+      "epoch": 0.3922484240018678,
+      "grad_norm": 0.8673185709111124,
+      "learning_rate": 0.00019873598452967338,
+      "loss": 2.3731,
+      "step": 420
+    },
+    {
+      "epoch": 0.4015876721923885,
+      "grad_norm": 0.22424350669971718,
+      "learning_rate": 0.0001986105777975613,
+      "loss": 2.6195,
+      "step": 430
+    },
+    {
+      "epoch": 0.4109269203829092,
+      "grad_norm": 0.307418135168262,
+      "learning_rate": 0.00019847928416977126,
+      "loss": 2.3624,
+      "step": 440
+    },
+    {
+      "epoch": 0.42026616857342985,
+      "grad_norm": 0.07944722668080402,
+      "learning_rate": 0.00019834211148432536,
+      "loss": 2.2799,
+      "step": 450
+    },
+    {
+      "epoch": 0.4296054167639505,
+      "grad_norm": 0.18146933758664588,
+      "learning_rate": 0.00019819906793021614,
+      "loss": 2.2177,
+      "step": 460
+    },
+    {
+      "epoch": 0.43894466495447115,
+      "grad_norm": 0.07035825837333018,
+      "learning_rate": 0.0001980501620469178,
+      "loss": 2.1767,
+      "step": 470
+    },
+    {
+      "epoch": 0.44828391314499183,
+      "grad_norm": 0.04596186944454228,
+      "learning_rate": 0.0001978954027238763,
+      "loss": 2.1598,
+      "step": 480
+    },
+    {
+      "epoch": 0.4576231613355125,
+      "grad_norm": 0.041342347745088055,
+      "learning_rate": 0.0001977347991999786,
+      "loss": 2.131,
+      "step": 490
+    },
+    {
+      "epoch": 0.46696240952603313,
+      "grad_norm": 0.04172063219841485,
+      "learning_rate": 0.00019756836106300137,
+      "loss": 2.1231,
+      "step": 500
+    },
+    {
+      "epoch": 0.4763016577165538,
+      "grad_norm": 0.03373646457711144,
+      "learning_rate": 0.00019739609824903843,
+      "loss": 2.1146,
+      "step": 510
+    },
+    {
+      "epoch": 0.4856409059070745,
+      "grad_norm": 0.03736871030676605,
+      "learning_rate": 0.00019721802104190748,
+      "loss": 2.1003,
+      "step": 520
+    },
+    {
+      "epoch": 0.49498015409759516,
+      "grad_norm": 0.033931028038211034,
+      "learning_rate": 0.00019703414007253645,
+      "loss": 2.0983,
+      "step": 530
+    },
+    {
+      "epoch": 0.5043194022881158,
+      "grad_norm": 0.03790055446070549,
+      "learning_rate": 0.00019684446631832868,
+      "loss": 2.092,
+      "step": 540
+    },
+    {
+      "epoch": 0.5136586504786365,
+      "grad_norm": 0.030956192803893078,
+      "learning_rate": 0.00019664901110250758,
+      "loss": 2.0807,
+      "step": 550
+    },
+    {
+      "epoch": 0.5229978986691571,
+      "grad_norm": 0.03542530209935129,
+      "learning_rate": 0.00019644778609344068,
+      "loss": 2.0773,
+      "step": 560
+    },
+    {
+      "epoch": 0.5323371468596778,
+      "grad_norm": 0.040947757568902336,
+      "learning_rate": 0.00019624080330394306,
+      "loss": 2.0649,
+      "step": 570
+    },
+    {
+      "epoch": 0.5416763950501985,
+      "grad_norm": 0.034273415973688146,
+      "learning_rate": 0.00019602807509056018,
+      "loss": 2.0479,
+      "step": 580
+    },
+    {
+      "epoch": 0.5510156432407192,
+      "grad_norm": 0.031427481498873144,
+      "learning_rate": 0.00019580961415283028,
+      "loss": 2.0563,
+      "step": 590
+    },
+    {
+      "epoch": 0.5603548914312397,
+      "grad_norm": 0.03141549752041532,
+      "learning_rate": 0.00019558543353252611,
+      "loss": 2.0503,
+      "step": 600
+    },
+    {
+      "epoch": 0.5696941396217604,
+      "grad_norm": 0.033012392726428204,
+      "learning_rate": 0.00019535554661287652,
+      "loss": 2.0389,
+      "step": 610
+    },
+    {
+      "epoch": 0.5790333878122811,
+      "grad_norm": 0.02913261992661444,
+      "learning_rate": 0.0001951199671177673,
+      "loss": 2.036,
+      "step": 620
+    },
+    {
+      "epoch": 0.5883726360028018,
+      "grad_norm": 0.030543903708435332,
+      "learning_rate": 0.00019487870911092214,
+      "loss": 2.0326,
+      "step": 630
+    },
+    {
+      "epoch": 0.5977118841933224,
+      "grad_norm": 0.03215005545393897,
+      "learning_rate": 0.00019463178699506277,
+      "loss": 2.0231,
+      "step": 640
+    },
+    {
+      "epoch": 0.6070511323838431,
+      "grad_norm": 0.03823630791937631,
+      "learning_rate": 0.00019437921551104933,
+      "loss": 2.0293,
+      "step": 650
+    },
+    {
+      "epoch": 0.6163903805743638,
+      "grad_norm": 0.03200103149471209,
+      "learning_rate": 0.00019412100973700038,
+      "loss": 2.017,
+      "step": 660
+    },
+    {
+      "epoch": 0.6257296287648845,
+      "grad_norm": 0.03841804153577787,
+      "learning_rate": 0.00019385718508739262,
+      "loss": 2.0135,
+      "step": 670
+    },
+    {
+      "epoch": 0.635068876955405,
+      "grad_norm": 0.03052396655271533,
+      "learning_rate": 0.0001935877573121407,
+      "loss": 2.0237,
+      "step": 680
+    },
+    {
+      "epoch": 0.6444081251459257,
+      "grad_norm": 0.033817837533771815,
+      "learning_rate": 0.00019331274249565717,
+      "loss": 2.0069,
+      "step": 690
+    },
+    {
+      "epoch": 0.6537473733364464,
+      "grad_norm": 0.028286999650643876,
+      "learning_rate": 0.00019303215705589194,
+      "loss": 2.0112,
+      "step": 700
+    },
+    {
+      "epoch": 0.6630866215269671,
+      "grad_norm": 0.030693948626257357,
+      "learning_rate": 0.00019274601774335243,
+      "loss": 2.0,
+      "step": 710
+    },
+    {
+      "epoch": 0.6724258697174877,
+      "grad_norm": 0.029984615635953022,
+      "learning_rate": 0.0001924543416401035,
+      "loss": 2.0028,
+      "step": 720
+    },
+    {
+      "epoch": 0.6817651179080084,
+      "grad_norm": 0.02915985673921391,
+      "learning_rate": 0.00019215714615874755,
+      "loss": 2.0031,
+      "step": 730
+    },
+    {
+      "epoch": 0.6911043660985291,
+      "grad_norm": 0.028305399777245336,
+      "learning_rate": 0.00019185444904138528,
+      "loss": 1.9924,
+      "step": 740
+    },
+    {
+      "epoch": 0.7004436142890498,
+      "grad_norm": 0.036720505429756495,
+      "learning_rate": 0.00019154626835855628,
+      "loss": 1.9981,
+      "step": 750
+    },
+    {
+      "epoch": 0.7097828624795703,
+      "grad_norm": 0.0287124048917296,
+      "learning_rate": 0.00019123262250816034,
+      "loss": 1.9868,
+      "step": 760
+    },
+    {
+      "epoch": 0.719122110670091,
+      "grad_norm": 0.03318092492837997,
+      "learning_rate": 0.00019091353021435915,
+      "loss": 1.9943,
+      "step": 770
+    },
+    {
+      "epoch": 0.7284613588606117,
+      "grad_norm": 0.054035272137015325,
+      "learning_rate": 0.00019058901052645844,
+      "loss": 1.9838,
+      "step": 780
+    },
+    {
+      "epoch": 0.7378006070511324,
+      "grad_norm": 0.03184392761983255,
+      "learning_rate": 0.00019025908281777078,
+      "loss": 1.982,
+      "step": 790
+    },
+    {
+      "epoch": 0.747139855241653,
+      "grad_norm": 0.029409948164434735,
+      "learning_rate": 0.00018992376678445908,
+      "loss": 1.9693,
+      "step": 800
+    },
+    {
+      "epoch": 0.7564791034321737,
+      "grad_norm": 0.029656963043919016,
+      "learning_rate": 0.00018958308244436064,
+      "loss": 1.9914,
+      "step": 810
+    },
+    {
+      "epoch": 0.7658183516226944,
+      "grad_norm": 0.030843610865326686,
+      "learning_rate": 0.00018923705013579233,
+      "loss": 1.9749,
+      "step": 820
+    },
+    {
+      "epoch": 0.7751575998132151,
+      "grad_norm": 0.0377584286045999,
+      "learning_rate": 0.00018888569051633613,
+      "loss": 1.9606,
+      "step": 830
+    },
+    {
+      "epoch": 0.7844968480037356,
+      "grad_norm": 0.03147699916274391,
+      "learning_rate": 0.00018852902456160616,
+      "loss": 1.9696,
+      "step": 840
+    },
+    {
+      "epoch": 0.7938360961942563,
+      "grad_norm": 0.029193153251471263,
+      "learning_rate": 0.0001881670735639963,
+      "loss": 1.9687,
+      "step": 850
+    },
+    {
+      "epoch": 0.803175344384777,
+      "grad_norm": 0.030201895228156087,
+      "learning_rate": 0.00018779985913140924,
+      "loss": 1.9678,
+      "step": 860
+    },
+    {
+      "epoch": 0.8125145925752977,
+      "grad_norm": 0.029272319485493213,
+      "learning_rate": 0.00018742740318596632,
+      "loss": 1.9697,
+      "step": 870
+    },
+    {
+      "epoch": 0.8218538407658184,
+      "grad_norm": 0.033740176465285654,
+      "learning_rate": 0.000187049727962699,
+      "loss": 1.9647,
+      "step": 880
+    },
+    {
+      "epoch": 0.831193088956339,
+      "grad_norm": 0.029539399251208593,
+      "learning_rate": 0.0001866668560082213,
+      "loss": 1.9627,
+      "step": 890
+    },
+    {
+      "epoch": 0.8405323371468597,
+      "grad_norm": 0.028666077337951026,
+      "learning_rate": 0.0001862788101793839,
+      "loss": 1.9529,
+      "step": 900
+    },
+    {
+      "epoch": 0.8498715853373804,
+      "grad_norm": 0.030280792492665805,
+      "learning_rate": 0.0001858856136419097,
+      "loss": 1.9536,
+      "step": 910
+    },
+    {
+      "epoch": 0.859210833527901,
+      "grad_norm": 0.03921536105057096,
+      "learning_rate": 0.0001854872898690106,
+      "loss": 1.9474,
+      "step": 920
+    },
+    {
+      "epoch": 0.8685500817184216,
+      "grad_norm": 0.030632523637038354,
+      "learning_rate": 0.0001850838626399865,
+      "loss": 1.9423,
+      "step": 930
+    },
+    {
+      "epoch": 0.8778893299089423,
+      "grad_norm": 0.04615147601979514,
+      "learning_rate": 0.00018467535603880548,
+      "loss": 1.946,
+      "step": 940
+    },
+    {
+      "epoch": 0.887228578099463,
+      "grad_norm": 0.028216236017006333,
+      "learning_rate": 0.00018426179445266616,
+      "loss": 1.9408,
+      "step": 950
+    },
+    {
+      "epoch": 0.8965678262899837,
+      "grad_norm": 0.0282407563402959,
+      "learning_rate": 0.00018384320257054177,
+      "loss": 1.9447,
+      "step": 960
+    },
+    {
+      "epoch": 0.9059070744805043,
+      "grad_norm": 0.029365880854687894,
+      "learning_rate": 0.0001834196053817062,
+      "loss": 1.9389,
+      "step": 970
+    },
+    {
+      "epoch": 0.915246322671025,
+      "grad_norm": 0.02855404439130719,
+      "learning_rate": 0.00018299102817424234,
+      "loss": 1.9425,
+      "step": 980
+    },
+    {
+      "epoch": 0.9245855708615457,
+      "grad_norm": 0.03238310984070135,
+      "learning_rate": 0.00018255749653353225,
+      "loss": 1.9392,
+      "step": 990
+    },
+    {
+      "epoch": 0.9339248190520663,
+      "grad_norm": 0.030115350805430388,
+      "learning_rate": 0.00018211903634072983,
+      "loss": 1.956,
+      "step": 1000
+    },
+    {
+      "epoch": 0.9432640672425869,
+      "grad_norm": 0.039090564032501135,
+      "learning_rate": 0.0001816756737712158,
+      "loss": 1.9358,
+      "step": 1010
+    },
+    {
+      "epoch": 0.9526033154331076,
+      "grad_norm": 0.028567954476327023,
+      "learning_rate": 0.000181227435293035,
+      "loss": 1.9342,
+      "step": 1020
+    },
+    {
+      "epoch": 0.9619425636236283,
+      "grad_norm": 0.028158653241284505,
+      "learning_rate": 0.00018077434766531624,
+      "loss": 1.9287,
+      "step": 1030
+    },
+    {
+      "epoch": 0.971281811814149,
+      "grad_norm": 0.02927795717651538,
+      "learning_rate": 0.00018031643793667504,
+      "loss": 1.9298,
+      "step": 1040
+    },
+    {
+      "epoch": 0.9806210600046696,
+      "grad_norm": 0.06298340584032344,
+      "learning_rate": 0.0001798537334435986,
+      "loss": 1.9303,
+      "step": 1050
+    },
+    {
+      "epoch": 0.9899603081951903,
+      "grad_norm": 0.03627278406983255,
+      "learning_rate": 0.00017938626180881407,
+      "loss": 1.9285,
+      "step": 1060
+    },
+    {
+      "epoch": 0.999299556385711,
+      "grad_norm": 0.03916823421329747,
+      "learning_rate": 0.00017891405093963938,
+      "loss": 1.9239,
+      "step": 1070
+    },
+    {
+      "epoch": 1.0087555451786132,
+      "grad_norm": 0.03383539251115568,
+      "learning_rate": 0.00017843712902631723,
+      "loss": 1.8855,
+      "step": 1080
+    },
+    {
+      "epoch": 1.0180947933691338,
+      "grad_norm": 0.02931021936560147,
+      "learning_rate": 0.00017795552454033224,
+      "loss": 1.8004,
+      "step": 1090
+    },
+    {
+      "epoch": 1.0274340415596543,
+      "grad_norm": 0.029173295095350292,
+      "learning_rate": 0.0001774692662327113,
+      "loss": 1.7912,
+      "step": 1100
+    },
+    {
+      "epoch": 1.0367732897501751,
+      "grad_norm": 0.035210065642974735,
+      "learning_rate": 0.000176978383132307,
+      "loss": 1.7902,
+      "step": 1110
+    },
+    {
+      "epoch": 1.0461125379406957,
+      "grad_norm": 0.029794447659573477,
+      "learning_rate": 0.00017648290454406475,
+      "loss": 1.8072,
+      "step": 1120
+    },
+    {
+      "epoch": 1.0554517861312165,
+      "grad_norm": 0.03194584938279939,
+      "learning_rate": 0.0001759828600472734,
+      "loss": 1.803,
+      "step": 1130
+    },
+    {
+      "epoch": 1.064791034321737,
+      "grad_norm": 0.031028415312581603,
+      "learning_rate": 0.00017547827949379924,
+      "loss": 1.7945,
+      "step": 1140
+    },
+    {
+      "epoch": 1.0741302825122578,
+      "grad_norm": 0.03313245828751237,
+      "learning_rate": 0.00017496919300630403,
+      "loss": 1.8139,
+      "step": 1150
+    },
+    {
+      "epoch": 1.0834695307027784,
+      "grad_norm": 0.03049431724979126,
+      "learning_rate": 0.00017445563097644664,
+      "loss": 1.8031,
+      "step": 1160
+    },
+    {
+      "epoch": 1.0928087788932992,
+      "grad_norm": 0.02880548705343715,
+      "learning_rate": 0.00017393762406306878,
+      "loss": 1.7974,
+      "step": 1170
+    },
+    {
+      "epoch": 1.1021480270838198,
+      "grad_norm": 0.03200427895977668,
+      "learning_rate": 0.00017341520319036469,
+      "loss": 1.7994,
+      "step": 1180
+    },
+    {
+      "epoch": 1.1114872752743403,
+      "grad_norm": 0.031293532815600045,
+      "learning_rate": 0.00017288839954603496,
+      "loss": 1.8073,
+      "step": 1190
+    },
+    {
+      "epoch": 1.1208265234648611,
+      "grad_norm": 0.032367211109345505,
+      "learning_rate": 0.00017235724457942468,
+      "loss": 1.7944,
+      "step": 1200
+    },
+    {
+      "epoch": 1.1301657716553817,
+      "grad_norm": 0.037783793923191374,
+      "learning_rate": 0.0001718217699996462,
+      "loss": 1.7948,
+      "step": 1210
+    },
+    {
+      "epoch": 1.1395050198459025,
+      "grad_norm": 0.02881083117349317,
+      "learning_rate": 0.00017128200777368567,
+      "loss": 1.8029,
+      "step": 1220
+    },
+    {
+      "epoch": 1.148844268036423,
+      "grad_norm": 0.032997408408175985,
+      "learning_rate": 0.00017073799012449524,
+      "loss": 1.7914,
+      "step": 1230
+    },
+    {
+      "epoch": 1.1581835162269438,
+      "grad_norm": 0.030263615801190885,
+      "learning_rate": 0.00017018974952906884,
+      "loss": 1.792,
+      "step": 1240
+    },
+    {
+      "epoch": 1.1675227644174644,
+      "grad_norm": 0.03286310429098484,
+      "learning_rate": 0.00016963731871650378,
+      "loss": 1.8149,
+      "step": 1250
+    },
+    {
+      "epoch": 1.1768620126079852,
+      "grad_norm": 0.031111621587597126,
+      "learning_rate": 0.00016908073066604663,
+      "loss": 1.8092,
+      "step": 1260
+    },
+    {
+      "epoch": 1.1862012607985057,
+      "grad_norm": 0.03372950172075241,
+      "learning_rate": 0.0001685200186051246,
+      "loss": 1.818,
+      "step": 1270
+    },
+    {
+      "epoch": 1.1955405089890263,
+      "grad_norm": 0.040782040436567434,
+      "learning_rate": 0.00016795521600736164,
+      "loss": 1.7999,
+      "step": 1280
+    },
+    {
+      "epoch": 1.204879757179547,
+      "grad_norm": 0.03451605168178924,
+      "learning_rate": 0.00016738635659058044,
+      "loss": 1.7945,
+      "step": 1290
+    },
+    {
+      "epoch": 1.2142190053700677,
+      "grad_norm": 0.03235681588882673,
+      "learning_rate": 0.00016681347431478933,
+      "loss": 1.8087,
+      "step": 1300
+    },
+    {
+      "epoch": 1.2235582535605884,
+      "grad_norm": 0.030750745605971932,
+      "learning_rate": 0.00016623660338015487,
+      "loss": 1.7995,
+      "step": 1310
+    },
+    {
+      "epoch": 1.232897501751109,
+      "grad_norm": 0.029444668665577274,
+      "learning_rate": 0.00016565577822496042,
+      "loss": 1.8025,
+      "step": 1320
+    },
+    {
+      "epoch": 1.2422367499416298,
+      "grad_norm": 0.038528856709584745,
+      "learning_rate": 0.00016507103352354996,
+      "loss": 1.7954,
+      "step": 1330
+    },
+    {
+      "epoch": 1.2515759981321504,
+      "grad_norm": 0.034217088004383035,
+      "learning_rate": 0.00016448240418425814,
+      "loss": 1.7962,
+      "step": 1340
+    },
+    {
+      "epoch": 1.260915246322671,
+      "grad_norm": 0.030205405393195585,
+      "learning_rate": 0.00016388992534732645,
+      "loss": 1.7973,
+      "step": 1350
+    },
+    {
+      "epoch": 1.2702544945131917,
+      "grad_norm": 0.029082218516562994,
+      "learning_rate": 0.00016329363238280528,
+      "loss": 1.796,
+      "step": 1360
+    },
+    {
+      "epoch": 1.2795937427037123,
+      "grad_norm": 0.029003887688766505,
+      "learning_rate": 0.00016269356088844238,
+      "loss": 1.7946,
+      "step": 1370
+    },
+    {
+      "epoch": 1.288932990894233,
+      "grad_norm": 0.03341157363649238,
+      "learning_rate": 0.00016208974668755779,
+      "loss": 1.7972,
+      "step": 1380
+    },
+    {
+      "epoch": 1.2982722390847536,
+      "grad_norm": 0.030614480844663026,
+      "learning_rate": 0.00016148222582690517,
+      "loss": 1.7973,
+      "step": 1390
+    },
+    {
+      "epoch": 1.3076114872752744,
+      "grad_norm": 0.029741346740467405,
+      "learning_rate": 0.00016087103457452,
+      "loss": 1.8076,
+      "step": 1400
+    },
+    {
+      "epoch": 1.316950735465795,
+      "grad_norm": 0.029569313554185597,
+      "learning_rate": 0.00016025620941755424,
+      "loss": 1.8043,
+      "step": 1410
+    },
+    {
+      "epoch": 1.3262899836563156,
+      "grad_norm": 0.02947637404374054,
+      "learning_rate": 0.0001596377870600983,
+      "loss": 1.797,
+      "step": 1420
+    },
+    {
+      "epoch": 1.3356292318468364,
+      "grad_norm": 0.031005062093959545,
+      "learning_rate": 0.00015901580442098968,
+      "loss": 1.8086,
+      "step": 1430
+    },
+    {
+      "epoch": 1.344968480037357,
+      "grad_norm": 0.029493792984873927,
+      "learning_rate": 0.00015839029863160922,
+      "loss": 1.8026,
+      "step": 1440
+    },
+    {
+      "epoch": 1.3543077282278777,
+      "grad_norm": 0.0288068155951218,
+      "learning_rate": 0.0001577613070336641,
+      "loss": 1.7951,
+      "step": 1450
+    },
+    {
+      "epoch": 1.3636469764183983,
+      "grad_norm": 0.03380404824627639,
+      "learning_rate": 0.00015712886717695885,
+      "loss": 1.7938,
+      "step": 1460
+    },
+    {
+      "epoch": 1.372986224608919,
+      "grad_norm": 0.039744706189693335,
+      "learning_rate": 0.0001564930168171536,
+      "loss": 1.8016,
+      "step": 1470
+    },
+    {
+      "epoch": 1.3823254727994396,
+      "grad_norm": 0.030565530594285437,
+      "learning_rate": 0.00015585379391351012,
+      "loss": 1.7984,
+      "step": 1480
+    },
+    {
+      "epoch": 1.3916647209899602,
+      "grad_norm": 0.04009392805554255,
+      "learning_rate": 0.00015521123662662567,
+      "loss": 1.7999,
+      "step": 1490
+    },
+    {
+      "epoch": 1.401003969180481,
+      "grad_norm": 0.03516196009586836,
+      "learning_rate": 0.000154565383316155,
+      "loss": 1.7979,
+      "step": 1500
+    },
+    {
+      "epoch": 1.4103432173710018,
+      "grad_norm": 0.03534161399054556,
+      "learning_rate": 0.0001539162725385202,
+      "loss": 1.8057,
+      "step": 1510
+    },
+    {
+      "epoch": 1.4196824655615223,
+      "grad_norm": 0.028488879438601067,
+      "learning_rate": 0.000153263943044609,
+      "loss": 1.792,
+      "step": 1520
+    },
+    {
+      "epoch": 1.429021713752043,
+      "grad_norm": 0.03125154490954804,
+      "learning_rate": 0.00015260843377746147,
+      "loss": 1.8008,
+      "step": 1530
+    },
+    {
+      "epoch": 1.4383609619425637,
+      "grad_norm": 0.030194357488801882,
+      "learning_rate": 0.00015194978386994507,
+      "loss": 1.7948,
+      "step": 1540
+    },
+    {
+      "epoch": 1.4477002101330843,
+      "grad_norm": 0.03049246845786265,
+      "learning_rate": 0.00015128803264241852,
+      "loss": 1.7967,
+      "step": 1550
+    },
+    {
+      "epoch": 1.4570394583236048,
+      "grad_norm": 0.030497211097258083,
+      "learning_rate": 0.0001506232196003844,
+      "loss": 1.7894,
+      "step": 1560
+    },
+    {
+      "epoch": 1.4663787065141256,
+      "grad_norm": 0.028748806119737205,
+      "learning_rate": 0.00014995538443213094,
+      "loss": 1.806,
+      "step": 1570
+    },
+    {
+      "epoch": 1.4757179547046464,
+      "grad_norm": 0.036423750322912396,
+      "learning_rate": 0.00014928456700636237,
+      "loss": 1.7995,
+      "step": 1580
+    },
+    {
+      "epoch": 1.485057202895167,
+      "grad_norm": 0.039101516109204065,
+      "learning_rate": 0.00014861080736981906,
+      "loss": 1.8028,
+      "step": 1590
+    },
+    {
+      "epoch": 1.4943964510856875,
+      "grad_norm": 0.031368399541673815,
+      "learning_rate": 0.00014793414574488663,
+      "loss": 1.8005,
+      "step": 1600
+    },
+    {
+      "epoch": 1.5037356992762083,
+      "grad_norm": 0.029788484702512056,
+      "learning_rate": 0.00014725462252719495,
+      "loss": 1.7963,
+      "step": 1610
+    },
+    {
+      "epoch": 1.5130749474667289,
+      "grad_norm": 0.029719041811636312,
+      "learning_rate": 0.00014657227828320635,
+      "loss": 1.7957,
+      "step": 1620
+    },
+    {
+      "epoch": 1.5224141956572494,
+      "grad_norm": 0.02820041575417432,
+      "learning_rate": 0.00014588715374779407,
+      "loss": 1.7986,
+      "step": 1630
+    },
+    {
+      "epoch": 1.5317534438477702,
+      "grad_norm": 1.8322544897261024,
+      "learning_rate": 0.0001451992898218102,
+      "loss": 2.016,
+      "step": 1640
+    },
+    {
+      "epoch": 1.541092692038291,
+      "grad_norm": 0.8676995793107466,
+      "learning_rate": 0.0001445087275696443,
+      "loss": 2.1095,
+      "step": 1650
+    },
+    {
+      "epoch": 1.5504319402288116,
+      "grad_norm": 0.08688193434631736,
+      "learning_rate": 0.00014381550821677155,
+      "loss": 2.0497,
+      "step": 1660
+    },
+    {
+      "epoch": 1.5597711884193322,
+      "grad_norm": 0.18321381224589608,
+      "learning_rate": 0.0001431196731472921,
+      "loss": 1.9023,
+      "step": 1670
+    },
+    {
+      "epoch": 1.569110436609853,
+      "grad_norm": 0.058214343698110564,
+      "learning_rate": 0.00014242126390145998,
+      "loss": 1.8566,
+      "step": 1680
+    },
+    {
+      "epoch": 1.5784496848003737,
+      "grad_norm": 0.03965578147557666,
+      "learning_rate": 0.0001417203221732036,
+      "loss": 1.8206,
+      "step": 1690
+    },
+    {
+      "epoch": 1.587788932990894,
+      "grad_norm": 0.03131802880017099,
+      "learning_rate": 0.00014101688980763658,
+      "loss": 1.8272,
+      "step": 1700
+    },
+    {
+      "epoch": 1.5971281811814149,
+      "grad_norm": 0.02788722706683908,
+      "learning_rate": 0.00014031100879855968,
+      "loss": 1.8145,
+      "step": 1710
+    },
+    {
+      "epoch": 1.6064674293719357,
+      "grad_norm": 0.02781057130092059,
+      "learning_rate": 0.00013960272128595372,
+      "loss": 1.8122,
+      "step": 1720
+    },
+    {
+      "epoch": 1.6158066775624562,
+      "grad_norm": 0.032067383734154756,
+      "learning_rate": 0.00013889206955346403,
+      "loss": 1.8064,
+      "step": 1730
+    },
+    {
+      "epoch": 1.6251459257529768,
+      "grad_norm": 0.03448807133884858,
+      "learning_rate": 0.00013817909602587613,
+      "loss": 1.8128,
+      "step": 1740
+    },
+    {
+      "epoch": 1.6344851739434976,
+      "grad_norm": 0.04127395288271696,
+      "learning_rate": 0.00013746384326658305,
+      "loss": 1.8041,
+      "step": 1750
+    },
+    {
+      "epoch": 1.6438244221340184,
+      "grad_norm": 0.030243192787820814,
+      "learning_rate": 0.00013674635397504427,
+      "loss": 1.803,
+      "step": 1760
+    },
+    {
+      "epoch": 1.653163670324539,
+      "grad_norm": 0.032183525625428915,
+      "learning_rate": 0.00013602667098423687,
+      "loss": 1.8092,
+      "step": 1770
+    },
+    {
+      "epoch": 1.6625029185150595,
+      "grad_norm": 0.030391044676815377,
+      "learning_rate": 0.00013530483725809818,
+      "loss": 1.8039,
+      "step": 1780
+    },
+    {
+      "epoch": 1.6718421667055803,
+      "grad_norm": 0.03204756818238517,
+      "learning_rate": 0.000134580895888961,
+      "loss": 1.8017,
+      "step": 1790
+    },
+    {
+      "epoch": 1.6811814148961008,
+      "grad_norm": 0.031117112662107062,
+      "learning_rate": 0.00013385489009498124,
+      "loss": 1.8017,
+      "step": 1800
+    },
+    {
+      "epoch": 1.6905206630866214,
+      "grad_norm": 0.028389293445401805,
+      "learning_rate": 0.00013312686321755761,
+      "loss": 1.811,
+      "step": 1810
+    },
+    {
+      "epoch": 1.6998599112771422,
+      "grad_norm": 4.908185097372493,
+      "learning_rate": 0.0001323968587187443,
+      "loss": 2.194,
+      "step": 1820
+    },
+    {
+      "epoch": 1.709199159467663,
+      "grad_norm": 0.2849452041194025,
+      "learning_rate": 0.00013166492017865637,
+      "loss": 2.0785,
+      "step": 1830
+    },
+    {
+      "epoch": 1.7185384076581836,
+      "grad_norm": 0.10989252058989733,
+      "learning_rate": 0.0001309310912928682,
+      "loss": 1.986,
+      "step": 1840
+    },
+    {
+      "epoch": 1.7278776558487041,
+      "grad_norm": 0.057475612656740484,
+      "learning_rate": 0.00013019541586980463,
+      "loss": 1.8614,
+      "step": 1850
+    },
+    {
+      "epoch": 1.737216904039225,
+      "grad_norm": 0.034908181734207726,
+      "learning_rate": 0.000129457937828126,
+      "loss": 1.8326,
+      "step": 1860
+    },
+    {
+      "epoch": 1.7465561522297455,
+      "grad_norm": 0.02892836681897248,
+      "learning_rate": 0.00012871870119410614,
+      "loss": 1.8243,
+      "step": 1870
+    },
+    {
+      "epoch": 1.755895400420266,
+      "grad_norm": 0.03311636981729384,
+      "learning_rate": 0.00012797775009900397,
+      "loss": 1.8183,
+      "step": 1880
+    },
+    {
+      "epoch": 1.7652346486107868,
+      "grad_norm": 0.03151917952358458,
+      "learning_rate": 0.00012723512877642904,
+      "loss": 1.8034,
+      "step": 1890
+    },
+    {
+      "epoch": 1.7745738968013076,
+      "grad_norm": 0.028109921832296925,
+      "learning_rate": 0.000126490881559701,
+      "loss": 1.8129,
+      "step": 1900
+    },
+    {
+      "epoch": 1.7839131449918282,
+      "grad_norm": 0.030350462454962698,
+      "learning_rate": 0.00012574505287920259,
+      "loss": 1.8003,
+      "step": 1910
+    },
+    {
+      "epoch": 1.7932523931823487,
+      "grad_norm": 0.03131380630103849,
+      "learning_rate": 0.00012499768725972754,
+      "loss": 1.814,
+      "step": 1920
+    },
+    {
+      "epoch": 1.8025916413728695,
+      "grad_norm": 0.029450198273050322,
+      "learning_rate": 0.00012424882931782243,
+      "loss": 1.7998,
+      "step": 1930
+    },
+    {
+      "epoch": 1.81193088956339,
+      "grad_norm": 0.0310524261453681,
+      "learning_rate": 0.0001234985237591231,
+      "loss": 1.8078,
+      "step": 1940
+    },
+    {
+      "epoch": 1.8212701377539107,
+      "grad_norm": 0.029362038478982822,
+      "learning_rate": 0.00012274681537568585,
+      "loss": 1.8014,
+      "step": 1950
+    },
+    {
+      "epoch": 1.8306093859444315,
+      "grad_norm": 0.027166816460226118,
+      "learning_rate": 0.00012199374904331337,
+      "loss": 1.8021,
+      "step": 1960
+    },
+    {
+      "epoch": 1.8399486341349522,
+      "grad_norm": 0.03680509283276228,
+      "learning_rate": 0.00012123936971887578,
+      "loss": 1.7973,
+      "step": 1970
+    },
+    {
+      "epoch": 1.8492878823254728,
+      "grad_norm": 0.03135174840346185,
+      "learning_rate": 0.0001204837224376267,
+      "loss": 1.7874,
+      "step": 1980
+    },
+    {
+      "epoch": 1.8586271305159934,
+      "grad_norm": 0.02644533256389969,
+      "learning_rate": 0.0001197268523105148,
+      "loss": 1.798,
+      "step": 1990
+    },
+    {
+      "epoch": 1.8679663787065142,
+      "grad_norm": 0.02999453651649614,
+      "learning_rate": 0.00011896880452149077,
+      "loss": 1.7957,
+      "step": 2000
+    },
+    {
+      "epoch": 1.877305626897035,
+      "grad_norm": 0.026905209700322272,
+      "learning_rate": 0.00011820962432480985,
+      "loss": 1.793,
+      "step": 2010
+    },
+    {
+      "epoch": 1.8866448750875553,
+      "grad_norm": 0.027263640323285022,
+      "learning_rate": 0.00011744935704233005,
+      "loss": 1.7974,
+      "step": 2020
+    },
+    {
+      "epoch": 1.895984123278076,
+      "grad_norm": 0.030479226063932337,
+      "learning_rate": 0.00011668804806080693,
+      "loss": 1.7898,
+      "step": 2030
+    },
+    {
+      "epoch": 1.9053233714685969,
+      "grad_norm": 0.030129902025534238,
+      "learning_rate": 0.00011592574282918369,
+      "loss": 1.7856,
+      "step": 2040
+    },
+    {
+      "epoch": 1.9146626196591174,
+      "grad_norm": 0.027884976674153635,
+      "learning_rate": 0.00011516248685587814,
+      "loss": 1.7858,
+      "step": 2050
+    },
+    {
+      "epoch": 1.924001867849638,
+      "grad_norm": 0.02925266011156687,
+      "learning_rate": 0.00011439832570606586,
+      "loss": 1.7876,
+      "step": 2060
+    },
+    {
+      "epoch": 1.9333411160401588,
+      "grad_norm": 0.028472914828616754,
+      "learning_rate": 0.00011363330499895997,
+      "loss": 1.7834,
+      "step": 2070
+    },
+    {
+      "epoch": 1.9426803642306796,
+      "grad_norm": 0.025877740032137875,
+      "learning_rate": 0.00011286747040508789,
+      "loss": 1.7955,
+      "step": 2080
+    },
+    {
+      "epoch": 1.9520196124212001,
+      "grad_norm": 0.02605295620697312,
+      "learning_rate": 0.0001121008676435648,
+      "loss": 1.7877,
+      "step": 2090
+    },
+    {
+      "epoch": 1.9613588606117207,
+      "grad_norm": 0.026887649929567867,
+      "learning_rate": 0.00011133354247936423,
+      "loss": 1.773,
+      "step": 2100
+    },
+    {
+      "epoch": 1.9706981088022415,
+      "grad_norm": 0.027982045915154026,
+      "learning_rate": 0.00011056554072058596,
+      "loss": 1.7762,
+      "step": 2110
+    },
+    {
+      "epoch": 1.980037356992762,
+      "grad_norm": 0.028077027059053006,
+      "learning_rate": 0.0001097969082157215,
+      "loss": 1.7963,
+      "step": 2120
+    },
+    {
+      "epoch": 1.9893766051832826,
+      "grad_norm": 0.027877078975954036,
+      "learning_rate": 0.00010902769085091686,
+      "loss": 1.7787,
+      "step": 2130
+    },
+    {
+      "epoch": 1.9987158533738034,
+      "grad_norm": 0.026120077388738373,
+      "learning_rate": 0.00010825793454723325,
+      "loss": 1.7842,
+      "step": 2140
+    },
+    {
+      "epoch": 2.0081718421667056,
+      "grad_norm": 0.044810283079268924,
+      "learning_rate": 0.00010748768525790569,
+      "loss": 1.6591,
+      "step": 2150
+    },
+    {
+      "epoch": 2.0175110903572264,
+      "grad_norm": 0.03431848280739808,
+      "learning_rate": 0.00010671698896559968,
+      "loss": 1.5599,
+      "step": 2160
+    },
+    {
+      "epoch": 2.0268503385477468,
+      "grad_norm": 0.04234332973849956,
+      "learning_rate": 0.00010594589167966606,
+      "loss": 1.5494,
+      "step": 2170
+    },
+    {
+      "epoch": 2.0361895867382676,
+      "grad_norm": 0.03260321438171042,
+      "learning_rate": 0.00010517443943339438,
+      "loss": 1.5473,
+      "step": 2180
+    },
+    {
+      "epoch": 2.0455288349287883,
+      "grad_norm": 0.034116901745609114,
+      "learning_rate": 0.00010440267828126478,
+      "loss": 1.5464,
+      "step": 2190
+    },
+    {
+      "epoch": 2.0548680831193087,
+      "grad_norm": 0.030992757239375807,
+      "learning_rate": 0.00010363065429619858,
+      "loss": 1.5514,
+      "step": 2200
+    },
+    {
+      "epoch": 2.0642073313098295,
+      "grad_norm": 0.03365516197786113,
+      "learning_rate": 0.0001028584135668077,
+      "loss": 1.5493,
+      "step": 2210
+    },
+    {
+      "epoch": 2.0735465795003503,
+      "grad_norm": 0.033293307482261586,
+      "learning_rate": 0.00010208600219464355,
+      "loss": 1.5426,
+      "step": 2220
+    },
+    {
+      "epoch": 2.082885827690871,
+      "grad_norm": 0.03653097737467338,
+      "learning_rate": 0.00010131346629144451,
+      "loss": 1.5471,
+      "step": 2230
+    },
+    {
+      "epoch": 2.0922250758813914,
+      "grad_norm": 0.03390291511697895,
+      "learning_rate": 0.0001005408519763833,
+      "loss": 1.5568,
+      "step": 2240
+    },
+    {
+      "epoch": 2.101564324071912,
+      "grad_norm": 0.03192694661852283,
+      "learning_rate": 9.976820537331374e-05,
+      "loss": 1.5452,
+      "step": 2250
+    },
+    {
+      "epoch": 2.110903572262433,
+      "grad_norm": 0.03561740193515691,
+      "learning_rate": 9.899557260801707e-05,
+      "loss": 1.546,
+      "step": 2260
+    },
+    {
+      "epoch": 2.1202428204529538,
+      "grad_norm": 0.029803732953068658,
+      "learning_rate": 9.822299980544862e-05,
+      "loss": 1.5533,
+      "step": 2270
+    },
+    {
+      "epoch": 2.129582068643474,
+      "grad_norm": 0.031232417271289125,
+      "learning_rate": 9.745053308698392e-05,
+      "loss": 1.5469,
+      "step": 2280
+    },
+    {
+      "epoch": 2.138921316833995,
+      "grad_norm": 0.032434793780181034,
+      "learning_rate": 9.667821856766548e-05,
+      "loss": 1.5514,
+      "step": 2290
+    },
+    {
+      "epoch": 2.1482605650245157,
+      "grad_norm": 0.03579370906405582,
+      "learning_rate": 9.590610235344972e-05,
+      "loss": 1.5577,
+      "step": 2300
+    },
+    {
+      "epoch": 2.157599813215036,
+      "grad_norm": 0.029662202478648328,
+      "learning_rate": 9.51342305384546e-05,
+      "loss": 1.5543,
+      "step": 2310
+    },
+    {
+      "epoch": 2.166939061405557,
+      "grad_norm": 0.03178715913934592,
+      "learning_rate": 9.436264920220781e-05,
+      "loss": 1.5579,
+      "step": 2320
+    },
+    {
+      "epoch": 2.1762783095960776,
+      "grad_norm": 0.03384008887051677,
+      "learning_rate": 9.359140440689601e-05,
+      "loss": 1.5595,
+      "step": 2330
+    },
+    {
+      "epoch": 2.1856175577865984,
+      "grad_norm": 0.03316450664408166,
+      "learning_rate": 9.282054219461475e-05,
+      "loss": 1.5556,
+      "step": 2340
+    },
+    {
+      "epoch": 2.1949568059771187,
+      "grad_norm": 0.032176305552558876,
+      "learning_rate": 9.205010858462007e-05,
+      "loss": 1.5638,
+      "step": 2350
+    },
+    {
+      "epoch": 2.2042960541676395,
+      "grad_norm": 0.031134335256756362,
+      "learning_rate": 9.128014957058109e-05,
+      "loss": 1.5629,
+      "step": 2360
+    },
+    {
+      "epoch": 2.2136353023581603,
+      "grad_norm": 0.032205851810441756,
+      "learning_rate": 9.051071111783436e-05,
+      "loss": 1.5613,
+      "step": 2370
+    },
+    {
+      "epoch": 2.2229745505486807,
+      "grad_norm": 0.029048245476020442,
+      "learning_rate": 8.974183916063968e-05,
+      "loss": 1.5594,
+      "step": 2380
+    },
+    {
+      "epoch": 2.2323137987392014,
+      "grad_norm": 0.03183166054573621,
+      "learning_rate": 8.897357959943795e-05,
+      "loss": 1.5606,
+      "step": 2390
+    },
+    {
+      "epoch": 2.2416530469297222,
+      "grad_norm": 0.032309922117136916,
+      "learning_rate": 8.820597829811109e-05,
+      "loss": 1.5524,
+      "step": 2400
+    },
+    {
+      "epoch": 2.250992295120243,
+      "grad_norm": 0.03598922231958808,
+      "learning_rate": 8.743908108124388e-05,
+      "loss": 1.5604,
+      "step": 2410
+    },
+    {
+      "epoch": 2.2603315433107634,
+      "grad_norm": 0.03046424735786346,
+      "learning_rate": 8.667293373138835e-05,
+      "loss": 1.5598,
+      "step": 2420
+    },
+    {
+      "epoch": 2.269670791501284,
+      "grad_norm": 0.030995453538377543,
+      "learning_rate": 8.59075819863307e-05,
+      "loss": 1.5652,
+      "step": 2430
+    },
+    {
+      "epoch": 2.279010039691805,
+      "grad_norm": 0.029309020623010097,
+      "learning_rate": 8.514307153636077e-05,
+      "loss": 1.5651,
+      "step": 2440
+    },
+    {
+      "epoch": 2.2883492878823253,
+      "grad_norm": 0.03158721106736763,
+      "learning_rate": 8.437944802154434e-05,
+      "loss": 1.5581,
+      "step": 2450
+    },
+    {
+      "epoch": 2.297688536072846,
+      "grad_norm": 0.03168229084049938,
+      "learning_rate": 8.361675702899871e-05,
+      "loss": 1.5671,
+      "step": 2460
+    },
+    {
+      "epoch": 2.307027784263367,
+      "grad_norm": 0.031335222148495136,
+      "learning_rate": 8.2855044090171e-05,
+      "loss": 1.5675,
+      "step": 2470
+    },
+    {
+      "epoch": 2.3163670324538876,
+      "grad_norm": 0.031073847325941303,
+      "learning_rate": 8.209435467811998e-05,
+      "loss": 1.5624,
+      "step": 2480
+    },
+    {
+      "epoch": 2.325706280644408,
+      "grad_norm": 0.030099100631045844,
+      "learning_rate": 8.133473420480161e-05,
+      "loss": 1.5606,
+      "step": 2490
+    },
+    {
+      "epoch": 2.3350455288349288,
+      "grad_norm": 0.034147632950361176,
+      "learning_rate": 8.057622801835788e-05,
+      "loss": 1.5703,
+      "step": 2500
+    },
+    {
+      "epoch": 2.3443847770254496,
+      "grad_norm": 0.03051580784550685,
+      "learning_rate": 7.981888140040955e-05,
+      "loss": 1.5731,
+      "step": 2510
+    },
+    {
+      "epoch": 2.3537240252159704,
+      "grad_norm": 0.03068917065832597,
+      "learning_rate": 7.9062739563353e-05,
+      "loss": 1.5723,
+      "step": 2520
+    },
+    {
+      "epoch": 2.3630632734064907,
+      "grad_norm": 0.02899547641705554,
+      "learning_rate": 7.830784764766118e-05,
+      "loss": 1.5691,
+      "step": 2530
+    },
+    {
+      "epoch": 2.3724025215970115,
+      "grad_norm": 0.030965383166701443,
+      "learning_rate": 7.755425071918858e-05,
+      "loss": 1.5627,
+      "step": 2540
+    },
+    {
+      "epoch": 2.3817417697875323,
+      "grad_norm": 0.03252440018336625,
+      "learning_rate": 7.680199376648108e-05,
+      "loss": 1.5536,
+      "step": 2550
+    },
+    {
+      "epoch": 2.3910810179780526,
+      "grad_norm": 0.031720485449340044,
+      "learning_rate": 7.605112169809008e-05,
+      "loss": 1.5617,
+      "step": 2560
+    },
+    {
+      "epoch": 2.4004202661685734,
+      "grad_norm": 0.031796658969132544,
+      "learning_rate": 7.530167933989161e-05,
+      "loss": 1.5595,
+      "step": 2570
+    },
+    {
+      "epoch": 2.409759514359094,
+      "grad_norm": 0.03218288097429844,
+      "learning_rate": 7.45537114324102e-05,
+      "loss": 1.5628,
+      "step": 2580
+    },
+    {
+      "epoch": 2.4190987625496145,
+      "grad_norm": 0.0305713183075559,
+      "learning_rate": 7.380726262814814e-05,
+      "loss": 1.5717,
+      "step": 2590
+    },
+    {
+      "epoch": 2.4284380107401353,
+      "grad_norm": 0.06879342166341705,
+      "learning_rate": 7.30623774889195e-05,
+      "loss": 1.5726,
+      "step": 2600
+    },
+    {
+      "epoch": 2.437777258930656,
+      "grad_norm": 0.04101428600237338,
+      "learning_rate": 7.231910048319011e-05,
+      "loss": 1.5679,
+      "step": 2610
+    },
+    {
+      "epoch": 2.447116507121177,
+      "grad_norm": 0.031060002638443395,
+      "learning_rate": 7.157747598342274e-05,
+      "loss": 1.562,
+      "step": 2620
+    },
+    {
+      "epoch": 2.4564557553116972,
+      "grad_norm": 0.032302829437386466,
+      "learning_rate": 7.083754826342816e-05,
+      "loss": 1.5767,
+      "step": 2630
+    },
+    {
+      "epoch": 2.465795003502218,
+      "grad_norm": 0.03111462744196413,
+      "learning_rate": 7.009936149572205e-05,
+      "loss": 1.5672,
+      "step": 2640
+    },
+    {
+      "epoch": 2.475134251692739,
+      "grad_norm": 0.031134083521743777,
+      "learning_rate": 6.936295974888807e-05,
+      "loss": 1.5665,
+      "step": 2650
+    },
+    {
+      "epoch": 2.4844734998832596,
+      "grad_norm": 0.030961556373721985,
+      "learning_rate": 6.862838698494693e-05,
+      "loss": 1.5608,
+      "step": 2660
+    },
+    {
+      "epoch": 2.49381274807378,
+      "grad_norm": 0.03168121700432082,
+      "learning_rate": 6.789568705673183e-05,
+      "loss": 1.566,
+      "step": 2670
+    },
+    {
+      "epoch": 2.5031519962643007,
+      "grad_norm": 0.030850372541726772,
+      "learning_rate": 6.716490370527081e-05,
+      "loss": 1.5651,
+      "step": 2680
+    },
+    {
+      "epoch": 2.5124912444548215,
+      "grad_norm": 0.03076635908430861,
+      "learning_rate": 6.643608055717519e-05,
+      "loss": 1.5596,
+      "step": 2690
+    },
+    {
+      "epoch": 2.521830492645342,
+      "grad_norm": 0.031897253741779714,
+      "learning_rate": 6.570926112203528e-05,
+      "loss": 1.5716,
+      "step": 2700
+    },
+    {
+      "epoch": 2.5311697408358627,
+      "grad_norm": 0.03085546721246857,
+      "learning_rate": 6.498448878982291e-05,
+      "loss": 1.5647,
+      "step": 2710
+    },
+    {
+      "epoch": 2.5405089890263834,
+      "grad_norm": 0.03127518794548787,
+      "learning_rate": 6.426180682830107e-05,
+      "loss": 1.5573,
+      "step": 2720
+    },
+    {
+      "epoch": 2.549848237216904,
+      "grad_norm": 0.03196649247686066,
+      "learning_rate": 6.354125838044098e-05,
+      "loss": 1.5597,
+      "step": 2730
+    },
+    {
+      "epoch": 2.5591874854074246,
+      "grad_norm": 0.030359755432035333,
+      "learning_rate": 6.282288646184638e-05,
+      "loss": 1.5625,
+      "step": 2740
+    },
+    {
+      "epoch": 2.5685267335979454,
+      "grad_norm": 0.03030640438940187,
+      "learning_rate": 6.210673395818571e-05,
+      "loss": 1.5717,
+      "step": 2750
+    },
+    {
+      "epoch": 2.577865981788466,
+      "grad_norm": 0.032197470232298186,
+      "learning_rate": 6.139284362263185e-05,
+      "loss": 1.5663,
+      "step": 2760
+    },
+    {
+      "epoch": 2.587205229978987,
+      "grad_norm": 0.030983733397891462,
+      "learning_rate": 6.0681258073309756e-05,
+      "loss": 1.5657,
+      "step": 2770
+    },
+    {
+      "epoch": 2.5965444781695073,
+      "grad_norm": 0.030427577702286164,
+      "learning_rate": 5.9972019790752385e-05,
+      "loss": 1.5708,
+      "step": 2780
+    },
+    {
+      "epoch": 2.605883726360028,
+      "grad_norm": 0.032761226318855745,
+      "learning_rate": 5.9265171115364495e-05,
+      "loss": 1.5641,
+      "step": 2790
+    },
+    {
+      "epoch": 2.615222974550549,
+      "grad_norm": 0.0317533648622182,
+      "learning_rate": 5.856075424489511e-05,
+      "loss": 1.5613,
+      "step": 2800
+    },
+    {
+      "epoch": 2.624562222741069,
+      "grad_norm": 0.03251887796623927,
+      "learning_rate": 5.785881123191834e-05,
+      "loss": 1.5644,
+      "step": 2810
+    },
+    {
+      "epoch": 2.63390147093159,
+      "grad_norm": 0.033109556138937665,
+      "learning_rate": 5.7159383981322866e-05,
+      "loss": 1.5613,
+      "step": 2820
+    },
+    {
+      "epoch": 2.643240719122111,
+      "grad_norm": 0.03190494841434546,
+      "learning_rate": 5.646251424781044e-05,
+      "loss": 1.5597,
+      "step": 2830
+    },
+    {
+      "epoch": 2.652579967312631,
+      "grad_norm": 0.03131425467271855,
+      "learning_rate": 5.576824363340293e-05,
+      "loss": 1.5644,
+      "step": 2840
+    },
+    {
+      "epoch": 2.661919215503152,
+      "grad_norm": 0.032580122114687284,
+      "learning_rate": 5.507661358495904e-05,
+      "loss": 1.5651,
+      "step": 2850
+    },
+    {
+      "epoch": 2.6712584636936727,
+      "grad_norm": 0.031118294785578867,
+      "learning_rate": 5.4387665391699814e-05,
+      "loss": 1.5595,
+      "step": 2860
+    },
+    {
+      "epoch": 2.680597711884193,
+      "grad_norm": 0.030454480461742643,
+      "learning_rate": 5.370144018274371e-05,
+      "loss": 1.5607,
+      "step": 2870
+    },
+    {
+      "epoch": 2.689936960074714,
+      "grad_norm": 0.030828671799046907,
+      "learning_rate": 5.301797892465148e-05,
+      "loss": 1.5587,
+      "step": 2880
+    },
+    {
+      "epoch": 2.6992762082652346,
+      "grad_norm": 0.030541588338347705,
+      "learning_rate": 5.2337322418980204e-05,
+      "loss": 1.5698,
+      "step": 2890
+    },
+    {
+      "epoch": 2.7086154564557554,
+      "grad_norm": 0.030634684731728472,
+      "learning_rate": 5.16595112998477e-05,
+      "loss": 1.5628,
+      "step": 2900
+    },
+    {
+      "epoch": 2.717954704646276,
+      "grad_norm": 0.03045168911848818,
+      "learning_rate": 5.098458603150691e-05,
+      "loss": 1.5544,
+      "step": 2910
+    },
+    {
+      "epoch": 2.7272939528367965,
+      "grad_norm": 0.029966073421542574,
+      "learning_rate": 5.0312586905929816e-05,
+      "loss": 1.557,
+      "step": 2920
+    },
+    {
+      "epoch": 2.7366332010273173,
+      "grad_norm": 0.031203173596988765,
+      "learning_rate": 4.964355404040232e-05,
+      "loss": 1.5571,
+      "step": 2930
+    },
+    {
+      "epoch": 2.745972449217838,
+      "grad_norm": 0.02933060802902336,
+      "learning_rate": 4.897752737512944e-05,
+      "loss": 1.5518,
+      "step": 2940
+    },
+    {
+      "epoch": 2.7553116974083585,
+      "grad_norm": 0.03039237943996916,
+      "learning_rate": 4.8314546670850594e-05,
+      "loss": 1.5682,
+      "step": 2950
+    },
+    {
+      "epoch": 2.7646509455988793,
+      "grad_norm": 0.03077504425768828,
+      "learning_rate": 4.765465150646633e-05,
+      "loss": 1.5618,
+      "step": 2960
+    },
+    {
+      "epoch": 2.7739901937894,
+      "grad_norm": 0.02987361489256037,
+      "learning_rate": 4.699788127667517e-05,
+      "loss": 1.5657,
+      "step": 2970
+    },
+    {
+      "epoch": 2.7833294419799204,
+      "grad_norm": 0.030823634404763923,
+      "learning_rate": 4.634427518962209e-05,
+      "loss": 1.5611,
+      "step": 2980
+    },
+    {
+      "epoch": 2.792668690170441,
+      "grad_norm": 0.031136974955910973,
+      "learning_rate": 4.569387226455776e-05,
+      "loss": 1.558,
+      "step": 2990
+    },
+    {
+      "epoch": 2.802007938360962,
+      "grad_norm": 0.032345263568839946,
+      "learning_rate": 4.5046711329508997e-05,
+      "loss": 1.5567,
+      "step": 3000
+    },
+    {
+      "epoch": 2.8113471865514827,
+      "grad_norm": 0.03075590840408306,
+      "learning_rate": 4.440283101896112e-05,
+      "loss": 1.5509,
+      "step": 3010
+    },
+    {
+      "epoch": 2.8206864347420035,
+      "grad_norm": 0.03212380955577142,
+      "learning_rate": 4.376226977155118e-05,
+      "loss": 1.5549,
+      "step": 3020
+    },
+    {
+      "epoch": 2.830025682932524,
+      "grad_norm": 0.030448994102361984,
+      "learning_rate": 4.3125065827773535e-05,
+      "loss": 1.5582,
+      "step": 3030
+    },
+    {
+      "epoch": 2.8393649311230447,
+      "grad_norm": 0.02952309346631166,
+      "learning_rate": 4.249125722769679e-05,
+      "loss": 1.5621,
+      "step": 3040
+    },
+    {
+      "epoch": 2.8487041793135655,
+      "grad_norm": 0.029801760584304334,
+      "learning_rate": 4.18608818086928e-05,
+      "loss": 1.5566,
+      "step": 3050
+    },
+    {
+      "epoch": 2.858043427504086,
+      "grad_norm": 0.029682850572275397,
+      "learning_rate": 4.12339772031781e-05,
+      "loss": 1.563,
+      "step": 3060
+    },
+    {
+      "epoch": 2.8673826756946066,
+      "grad_norm": 0.0295800536252474,
+      "learning_rate": 4.061058083636702e-05,
+      "loss": 1.5478,
+      "step": 3070
+    },
+    {
+      "epoch": 2.8767219238851274,
+      "grad_norm": 0.029683305961870896,
+      "learning_rate": 3.999072992403756e-05,
+      "loss": 1.5628,
+      "step": 3080
+    },
+    {
+      "epoch": 2.8860611720756477,
+      "grad_norm": 0.029644699672258072,
+      "learning_rate": 3.93744614703098e-05,
+      "loss": 1.5541,
+      "step": 3090
+    },
+    {
+      "epoch": 2.8954004202661685,
+      "grad_norm": 0.030518095362273494,
+      "learning_rate": 3.876181226543668e-05,
+      "loss": 1.5702,
+      "step": 3100
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 4280,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 100,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.5100743120772923e+20,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}

l2-13b-ga/checkpoint-3100/zero_to_fp32.py

@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)

q2.5-eu-en/checkpoint-2400/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

q2.5-eu-en/checkpoint-2400/chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

q2.5-eu-en/checkpoint-2400/config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 896,
+  "initializer_range": 0.02,
+  "intermediate_size": 4864,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 24,
+  "model_type": "qwen2",
+  "num_attention_heads": 14,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": true,
+  "use_mrope": false,
+  "use_sliding_window": false,
+  "vocab_size": 151665
+}

q2.5-eu-en/checkpoint-2400/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "transformers_version": "4.52.4"
+}

q2.5-eu-en/checkpoint-2400/latest

@@ -0,0 +1 @@
+global_step2400

q2.5-eu-en/checkpoint-2400/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

q2.5-eu-en/checkpoint-2400/tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

q2.5-eu-en/checkpoint-2400/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

q2.5-eu-en/checkpoint-2700/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

q2.5-eu-en/checkpoint-2700/chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

q2.5-eu-en/checkpoint-2700/config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 896,
+  "initializer_range": 0.02,
+  "intermediate_size": 4864,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 24,
+  "model_type": "qwen2",
+  "num_attention_heads": 14,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": true,
+  "use_mrope": false,
+  "use_sliding_window": false,
+  "vocab_size": 151665
+}

q2.5-eu-en/checkpoint-2700/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "transformers_version": "4.52.4"
+}

q2.5-eu-en/checkpoint-2700/latest

@@ -0,0 +1 @@
+global_step2700

q2.5-eu-en/checkpoint-2700/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

q2.5-eu-en/checkpoint-2700/tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

q2.5-eu-en/checkpoint-2700/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

q2.5-eu-en/checkpoint-300/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

q2.5-eu-en/checkpoint-300/chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

q2.5-eu-en/checkpoint-300/config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 896,
+  "initializer_range": 0.02,
+  "intermediate_size": 4864,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 24,
+  "model_type": "qwen2",
+  "num_attention_heads": 14,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": true,
+  "use_mrope": false,
+  "use_sliding_window": false,
+  "vocab_size": 151665
+}

q2.5-eu-en/checkpoint-300/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "transformers_version": "4.52.4"
+}

q2.5-eu-en/checkpoint-300/latest

@@ -0,0 +1 @@
+global_step300

q2.5-eu-en/checkpoint-300/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

q2.5-eu-en/checkpoint-300/tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

q2.5-eu-en/checkpoint-300/trainer_state.json

@@ -0,0 +1,1091 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.1009633587143999,
+  "eval_steps": 500,
+  "global_step": 300,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.00033654452904799967,
+      "grad_norm": 4.331892967224121,
+      "learning_rate": 0.0,
+      "loss": 5.5158,
+      "step": 1
+    },
+    {
+      "epoch": 0.0006730890580959993,
+      "grad_norm": 4.539519786834717,
+      "learning_rate": 6.711409395973154e-07,
+      "loss": 5.5718,
+      "step": 2
+    },
+    {
+      "epoch": 0.0013461781161919987,
+      "grad_norm": 4.208465576171875,
+      "learning_rate": 2.013422818791946e-06,
+      "loss": 5.4905,
+      "step": 4
+    },
+    {
+      "epoch": 0.002019267174287998,
+      "grad_norm": 3.891338586807251,
+      "learning_rate": 3.3557046979865773e-06,
+      "loss": 5.4511,
+      "step": 6
+    },
+    {
+      "epoch": 0.0026923562323839974,
+      "grad_norm": 2.979590654373169,
+      "learning_rate": 4.697986577181209e-06,
+      "loss": 5.3311,
+      "step": 8
+    },
+    {
+      "epoch": 0.0033654452904799967,
+      "grad_norm": 3.1792373657226562,
+      "learning_rate": 6.04026845637584e-06,
+      "loss": 5.3782,
+      "step": 10
+    },
+    {
+      "epoch": 0.004038534348575996,
+      "grad_norm": 3.9909653663635254,
+      "learning_rate": 7.382550335570471e-06,
+      "loss": 5.2583,
+      "step": 12
+    },
+    {
+      "epoch": 0.004711623406671995,
+      "grad_norm": 3.31044340133667,
+      "learning_rate": 8.724832214765101e-06,
+      "loss": 5.3198,
+      "step": 14
+    },
+    {
+      "epoch": 0.005384712464767995,
+      "grad_norm": 3.3201687335968018,
+      "learning_rate": 1.006711409395973e-05,
+      "loss": 5.1138,
+      "step": 16
+    },
+    {
+      "epoch": 0.006057801522863994,
+      "grad_norm": 2.8034451007843018,
+      "learning_rate": 1.1409395973154363e-05,
+      "loss": 5.0022,
+      "step": 18
+    },
+    {
+      "epoch": 0.0067308905809599934,
+      "grad_norm": 2.044506072998047,
+      "learning_rate": 1.2751677852348994e-05,
+      "loss": 4.8879,
+      "step": 20
+    },
+    {
+      "epoch": 0.007403979639055993,
+      "grad_norm": 1.6194826364517212,
+      "learning_rate": 1.4093959731543624e-05,
+      "loss": 4.7436,
+      "step": 22
+    },
+    {
+      "epoch": 0.008077068697151992,
+      "grad_norm": 1.4383995532989502,
+      "learning_rate": 1.5436241610738255e-05,
+      "loss": 4.5798,
+      "step": 24
+    },
+    {
+      "epoch": 0.00875015775524799,
+      "grad_norm": 1.174633264541626,
+      "learning_rate": 1.6778523489932888e-05,
+      "loss": 4.4188,
+      "step": 26
+    },
+    {
+      "epoch": 0.00942324681334399,
+      "grad_norm": 1.1277130842208862,
+      "learning_rate": 1.8120805369127517e-05,
+      "loss": 4.4374,
+      "step": 28
+    },
+    {
+      "epoch": 0.01009633587143999,
+      "grad_norm": 1.0426617860794067,
+      "learning_rate": 1.946308724832215e-05,
+      "loss": 4.2195,
+      "step": 30
+    },
+    {
+      "epoch": 0.01076942492953599,
+      "grad_norm": 0.9180749654769897,
+      "learning_rate": 2.080536912751678e-05,
+      "loss": 4.1755,
+      "step": 32
+    },
+    {
+      "epoch": 0.011442513987631988,
+      "grad_norm": 1.0104376077651978,
+      "learning_rate": 2.2147651006711412e-05,
+      "loss": 4.0892,
+      "step": 34
+    },
+    {
+      "epoch": 0.012115603045727988,
+      "grad_norm": 0.9427777528762817,
+      "learning_rate": 2.348993288590604e-05,
+      "loss": 4.0498,
+      "step": 36
+    },
+    {
+      "epoch": 0.012788692103823987,
+      "grad_norm": 1.0120079517364502,
+      "learning_rate": 2.4832214765100674e-05,
+      "loss": 3.9471,
+      "step": 38
+    },
+    {
+      "epoch": 0.013461781161919987,
+      "grad_norm": 1.0842680931091309,
+      "learning_rate": 2.6174496644295304e-05,
+      "loss": 3.8882,
+      "step": 40
+    },
+    {
+      "epoch": 0.014134870220015985,
+      "grad_norm": 1.3515102863311768,
+      "learning_rate": 2.7516778523489933e-05,
+      "loss": 3.8274,
+      "step": 42
+    },
+    {
+      "epoch": 0.014807959278111986,
+      "grad_norm": 1.1003209352493286,
+      "learning_rate": 2.885906040268457e-05,
+      "loss": 3.7735,
+      "step": 44
+    },
+    {
+      "epoch": 0.015481048336207984,
+      "grad_norm": 0.9302487373352051,
+      "learning_rate": 3.02013422818792e-05,
+      "loss": 3.731,
+      "step": 46
+    },
+    {
+      "epoch": 0.016154137394303984,
+      "grad_norm": 1.0311543941497803,
+      "learning_rate": 3.1543624161073825e-05,
+      "loss": 3.6608,
+      "step": 48
+    },
+    {
+      "epoch": 0.016827226452399983,
+      "grad_norm": 1.0927435159683228,
+      "learning_rate": 3.288590604026846e-05,
+      "loss": 3.5962,
+      "step": 50
+    },
+    {
+      "epoch": 0.01750031551049598,
+      "grad_norm": 1.3255321979522705,
+      "learning_rate": 3.422818791946309e-05,
+      "loss": 3.5669,
+      "step": 52
+    },
+    {
+      "epoch": 0.018173404568591983,
+      "grad_norm": 0.9304305911064148,
+      "learning_rate": 3.557046979865772e-05,
+      "loss": 3.5456,
+      "step": 54
+    },
+    {
+      "epoch": 0.01884649362668798,
+      "grad_norm": 1.159280776977539,
+      "learning_rate": 3.6912751677852356e-05,
+      "loss": 3.5189,
+      "step": 56
+    },
+    {
+      "epoch": 0.01951958268478398,
+      "grad_norm": 0.9376134276390076,
+      "learning_rate": 3.8255033557046985e-05,
+      "loss": 3.432,
+      "step": 58
+    },
+    {
+      "epoch": 0.02019267174287998,
+      "grad_norm": 1.292802095413208,
+      "learning_rate": 3.959731543624161e-05,
+      "loss": 3.4564,
+      "step": 60
+    },
+    {
+      "epoch": 0.02086576080097598,
+      "grad_norm": 1.2383852005004883,
+      "learning_rate": 4.0939597315436244e-05,
+      "loss": 3.4194,
+      "step": 62
+    },
+    {
+      "epoch": 0.02153884985907198,
+      "grad_norm": 0.8546445369720459,
+      "learning_rate": 4.228187919463087e-05,
+      "loss": 3.3873,
+      "step": 64
+    },
+    {
+      "epoch": 0.022211938917167977,
+      "grad_norm": 1.4104743003845215,
+      "learning_rate": 4.36241610738255e-05,
+      "loss": 3.3647,
+      "step": 66
+    },
+    {
+      "epoch": 0.022885027975263976,
+      "grad_norm": 1.3548426628112793,
+      "learning_rate": 4.496644295302014e-05,
+      "loss": 3.3213,
+      "step": 68
+    },
+    {
+      "epoch": 0.023558117033359978,
+      "grad_norm": 1.1530455350875854,
+      "learning_rate": 4.630872483221477e-05,
+      "loss": 3.2789,
+      "step": 70
+    },
+    {
+      "epoch": 0.024231206091455976,
+      "grad_norm": 1.3092457056045532,
+      "learning_rate": 4.76510067114094e-05,
+      "loss": 3.2922,
+      "step": 72
+    },
+    {
+      "epoch": 0.024904295149551975,
+      "grad_norm": 1.577699065208435,
+      "learning_rate": 4.8993288590604034e-05,
+      "loss": 3.2308,
+      "step": 74
+    },
+    {
+      "epoch": 0.025577384207647973,
+      "grad_norm": 1.3348486423492432,
+      "learning_rate": 5.033557046979866e-05,
+      "loss": 3.2656,
+      "step": 76
+    },
+    {
+      "epoch": 0.026250473265743975,
+      "grad_norm": 1.1966625452041626,
+      "learning_rate": 5.167785234899329e-05,
+      "loss": 3.2703,
+      "step": 78
+    },
+    {
+      "epoch": 0.026923562323839974,
+      "grad_norm": 1.3125278949737549,
+      "learning_rate": 5.302013422818792e-05,
+      "loss": 3.2034,
+      "step": 80
+    },
+    {
+      "epoch": 0.027596651381935972,
+      "grad_norm": 1.1957862377166748,
+      "learning_rate": 5.436241610738255e-05,
+      "loss": 3.2476,
+      "step": 82
+    },
+    {
+      "epoch": 0.02826974044003197,
+      "grad_norm": 1.2177337408065796,
+      "learning_rate": 5.570469798657718e-05,
+      "loss": 3.2166,
+      "step": 84
+    },
+    {
+      "epoch": 0.028942829498127973,
+      "grad_norm": 0.8806389570236206,
+      "learning_rate": 5.704697986577181e-05,
+      "loss": 3.1722,
+      "step": 86
+    },
+    {
+      "epoch": 0.02961591855622397,
+      "grad_norm": 1.8180561065673828,
+      "learning_rate": 5.838926174496645e-05,
+      "loss": 3.1597,
+      "step": 88
+    },
+    {
+      "epoch": 0.03028900761431997,
+      "grad_norm": 1.1676297187805176,
+      "learning_rate": 5.973154362416108e-05,
+      "loss": 3.1829,
+      "step": 90
+    },
+    {
+      "epoch": 0.030962096672415968,
+      "grad_norm": 1.0163198709487915,
+      "learning_rate": 6.107382550335571e-05,
+      "loss": 3.1643,
+      "step": 92
+    },
+    {
+      "epoch": 0.03163518573051197,
+      "grad_norm": 1.0734015703201294,
+      "learning_rate": 6.241610738255034e-05,
+      "loss": 3.1692,
+      "step": 94
+    },
+    {
+      "epoch": 0.03230827478860797,
+      "grad_norm": 1.4144916534423828,
+      "learning_rate": 6.375838926174497e-05,
+      "loss": 3.1217,
+      "step": 96
+    },
+    {
+      "epoch": 0.03298136384670397,
+      "grad_norm": 1.5647915601730347,
+      "learning_rate": 6.51006711409396e-05,
+      "loss": 3.1324,
+      "step": 98
+    },
+    {
+      "epoch": 0.033654452904799965,
+      "grad_norm": 1.1999105215072632,
+      "learning_rate": 6.644295302013423e-05,
+      "loss": 3.1103,
+      "step": 100
+    },
+    {
+      "epoch": 0.03432754196289597,
+      "grad_norm": 1.512838363647461,
+      "learning_rate": 6.778523489932886e-05,
+      "loss": 3.1036,
+      "step": 102
+    },
+    {
+      "epoch": 0.03500063102099196,
+      "grad_norm": 1.1092990636825562,
+      "learning_rate": 6.912751677852349e-05,
+      "loss": 3.1244,
+      "step": 104
+    },
+    {
+      "epoch": 0.035673720079087964,
+      "grad_norm": 1.2763620615005493,
+      "learning_rate": 7.046979865771812e-05,
+      "loss": 3.0989,
+      "step": 106
+    },
+    {
+      "epoch": 0.036346809137183966,
+      "grad_norm": 1.2328648567199707,
+      "learning_rate": 7.181208053691275e-05,
+      "loss": 3.0506,
+      "step": 108
+    },
+    {
+      "epoch": 0.03701989819527996,
+      "grad_norm": 1.6039047241210938,
+      "learning_rate": 7.315436241610739e-05,
+      "loss": 3.0589,
+      "step": 110
+    },
+    {
+      "epoch": 0.03769298725337596,
+      "grad_norm": 1.7161307334899902,
+      "learning_rate": 7.449664429530202e-05,
+      "loss": 3.0296,
+      "step": 112
+    },
+    {
+      "epoch": 0.03836607631147196,
+      "grad_norm": 1.2628991603851318,
+      "learning_rate": 7.583892617449665e-05,
+      "loss": 3.063,
+      "step": 114
+    },
+    {
+      "epoch": 0.03903916536956796,
+      "grad_norm": 1.1414180994033813,
+      "learning_rate": 7.718120805369128e-05,
+      "loss": 3.0366,
+      "step": 116
+    },
+    {
+      "epoch": 0.03971225442766396,
+      "grad_norm": 1.5152932405471802,
+      "learning_rate": 7.852348993288591e-05,
+      "loss": 3.0241,
+      "step": 118
+    },
+    {
+      "epoch": 0.04038534348575996,
+      "grad_norm": 1.6119567155838013,
+      "learning_rate": 7.986577181208054e-05,
+      "loss": 3.0139,
+      "step": 120
+    },
+    {
+      "epoch": 0.04105843254385596,
+      "grad_norm": 1.6078091859817505,
+      "learning_rate": 8.120805369127518e-05,
+      "loss": 3.0288,
+      "step": 122
+    },
+    {
+      "epoch": 0.04173152160195196,
+      "grad_norm": 1.1092705726623535,
+      "learning_rate": 8.255033557046981e-05,
+      "loss": 2.9996,
+      "step": 124
+    },
+    {
+      "epoch": 0.042404610660047956,
+      "grad_norm": 1.2352242469787598,
+      "learning_rate": 8.389261744966444e-05,
+      "loss": 2.9872,
+      "step": 126
+    },
+    {
+      "epoch": 0.04307769971814396,
+      "grad_norm": 1.636400580406189,
+      "learning_rate": 8.523489932885907e-05,
+      "loss": 2.9814,
+      "step": 128
+    },
+    {
+      "epoch": 0.04375078877623995,
+      "grad_norm": 1.4877128601074219,
+      "learning_rate": 8.65771812080537e-05,
+      "loss": 2.9756,
+      "step": 130
+    },
+    {
+      "epoch": 0.044423877834335955,
+      "grad_norm": 1.2983709573745728,
+      "learning_rate": 8.791946308724833e-05,
+      "loss": 2.9756,
+      "step": 132
+    },
+    {
+      "epoch": 0.04509696689243196,
+      "grad_norm": 1.7350983619689941,
+      "learning_rate": 8.926174496644296e-05,
+      "loss": 2.9579,
+      "step": 134
+    },
+    {
+      "epoch": 0.04577005595052795,
+      "grad_norm": 0.978854775428772,
+      "learning_rate": 9.060402684563759e-05,
+      "loss": 2.9269,
+      "step": 136
+    },
+    {
+      "epoch": 0.046443145008623954,
+      "grad_norm": 1.362163782119751,
+      "learning_rate": 9.194630872483221e-05,
+      "loss": 2.99,
+      "step": 138
+    },
+    {
+      "epoch": 0.047116234066719956,
+      "grad_norm": 1.328202247619629,
+      "learning_rate": 9.328859060402684e-05,
+      "loss": 2.966,
+      "step": 140
+    },
+    {
+      "epoch": 0.04778932312481595,
+      "grad_norm": 1.2497445344924927,
+      "learning_rate": 9.463087248322147e-05,
+      "loss": 2.9254,
+      "step": 142
+    },
+    {
+      "epoch": 0.04846241218291195,
+      "grad_norm": 1.071092128753662,
+      "learning_rate": 9.59731543624161e-05,
+      "loss": 2.9597,
+      "step": 144
+    },
+    {
+      "epoch": 0.04913550124100795,
+      "grad_norm": 1.451729416847229,
+      "learning_rate": 9.731543624161075e-05,
+      "loss": 2.9605,
+      "step": 146
+    },
+    {
+      "epoch": 0.04980859029910395,
+      "grad_norm": 1.328731656074524,
+      "learning_rate": 9.865771812080538e-05,
+      "loss": 2.9493,
+      "step": 148
+    },
+    {
+      "epoch": 0.05048167935719995,
+      "grad_norm": 1.4675222635269165,
+      "learning_rate": 0.0001,
+      "loss": 2.9298,
+      "step": 150
+    },
+    {
+      "epoch": 0.05115476841529595,
+      "grad_norm": 1.208961009979248,
+      "learning_rate": 9.99998761551904e-05,
+      "loss": 2.9382,
+      "step": 152
+    },
+    {
+      "epoch": 0.05182785747339195,
+      "grad_norm": 1.0392056703567505,
+      "learning_rate": 9.999950462137508e-05,
+      "loss": 2.8829,
+      "step": 154
+    },
+    {
+      "epoch": 0.05250094653148795,
+      "grad_norm": 1.378711462020874,
+      "learning_rate": 9.999888540039458e-05,
+      "loss": 2.9041,
+      "step": 156
+    },
+    {
+      "epoch": 0.053174035589583946,
+      "grad_norm": 1.2687252759933472,
+      "learning_rate": 9.999801849531635e-05,
+      "loss": 2.9148,
+      "step": 158
+    },
+    {
+      "epoch": 0.05384712464767995,
+      "grad_norm": 1.2382102012634277,
+      "learning_rate": 9.999690391043487e-05,
+      "loss": 2.9107,
+      "step": 160
+    },
+    {
+      "epoch": 0.05452021370577594,
+      "grad_norm": 1.0215928554534912,
+      "learning_rate": 9.999554165127159e-05,
+      "loss": 2.9187,
+      "step": 162
+    },
+    {
+      "epoch": 0.055193302763871945,
+      "grad_norm": 1.2638540267944336,
+      "learning_rate": 9.99939317245748e-05,
+      "loss": 2.9261,
+      "step": 164
+    },
+    {
+      "epoch": 0.055866391821967946,
+      "grad_norm": 1.3826959133148193,
+      "learning_rate": 9.999207413831982e-05,
+      "loss": 2.8944,
+      "step": 166
+    },
+    {
+      "epoch": 0.05653948088006394,
+      "grad_norm": 1.0764875411987305,
+      "learning_rate": 9.998996890170867e-05,
+      "loss": 2.9031,
+      "step": 168
+    },
+    {
+      "epoch": 0.05721256993815994,
+      "grad_norm": 1.5163936614990234,
+      "learning_rate": 9.99876160251703e-05,
+      "loss": 2.8687,
+      "step": 170
+    },
+    {
+      "epoch": 0.057885658996255945,
+      "grad_norm": 1.5821291208267212,
+      "learning_rate": 9.998501552036037e-05,
+      "loss": 2.8828,
+      "step": 172
+    },
+    {
+      "epoch": 0.05855874805435194,
+      "grad_norm": 1.4572120904922485,
+      "learning_rate": 9.998216740016124e-05,
+      "loss": 2.8898,
+      "step": 174
+    },
+    {
+      "epoch": 0.05923183711244794,
+      "grad_norm": 1.2185218334197998,
+      "learning_rate": 9.99790716786819e-05,
+      "loss": 2.8456,
+      "step": 176
+    },
+    {
+      "epoch": 0.05990492617054394,
+      "grad_norm": 1.106446385383606,
+      "learning_rate": 9.99757283712579e-05,
+      "loss": 2.8793,
+      "step": 178
+    },
+    {
+      "epoch": 0.06057801522863994,
+      "grad_norm": 1.1448893547058105,
+      "learning_rate": 9.997213749445129e-05,
+      "loss": 2.8579,
+      "step": 180
+    },
+    {
+      "epoch": 0.06125110428673594,
+      "grad_norm": 1.1587834358215332,
+      "learning_rate": 9.996829906605056e-05,
+      "loss": 2.8839,
+      "step": 182
+    },
+    {
+      "epoch": 0.061924193344831936,
+      "grad_norm": 1.0969592332839966,
+      "learning_rate": 9.996421310507046e-05,
+      "loss": 2.8638,
+      "step": 184
+    },
+    {
+      "epoch": 0.06259728240292793,
+      "grad_norm": 0.9740116000175476,
+      "learning_rate": 9.9959879631752e-05,
+      "loss": 2.8455,
+      "step": 186
+    },
+    {
+      "epoch": 0.06327037146102393,
+      "grad_norm": 1.2307910919189453,
+      "learning_rate": 9.995529866756231e-05,
+      "loss": 2.8534,
+      "step": 188
+    },
+    {
+      "epoch": 0.06394346051911994,
+      "grad_norm": 1.5021939277648926,
+      "learning_rate": 9.995047023519452e-05,
+      "loss": 2.8469,
+      "step": 190
+    },
+    {
+      "epoch": 0.06461654957721594,
+      "grad_norm": 1.1044224500656128,
+      "learning_rate": 9.994539435856771e-05,
+      "loss": 2.8429,
+      "step": 192
+    },
+    {
+      "epoch": 0.06528963863531194,
+      "grad_norm": 1.4586883783340454,
+      "learning_rate": 9.99400710628267e-05,
+      "loss": 2.836,
+      "step": 194
+    },
+    {
+      "epoch": 0.06596272769340794,
+      "grad_norm": 1.2613426446914673,
+      "learning_rate": 9.993450037434199e-05,
+      "loss": 2.8243,
+      "step": 196
+    },
+    {
+      "epoch": 0.06663581675150393,
+      "grad_norm": 1.0347422361373901,
+      "learning_rate": 9.992868232070963e-05,
+      "loss": 2.7965,
+      "step": 198
+    },
+    {
+      "epoch": 0.06730890580959993,
+      "grad_norm": 2.1357574462890625,
+      "learning_rate": 9.992261693075103e-05,
+      "loss": 2.8486,
+      "step": 200
+    },
+    {
+      "epoch": 0.06798199486769593,
+      "grad_norm": 1.0357908010482788,
+      "learning_rate": 9.991630423451286e-05,
+      "loss": 2.8386,
+      "step": 202
+    },
+    {
+      "epoch": 0.06865508392579193,
+      "grad_norm": 1.1383159160614014,
+      "learning_rate": 9.990974426326696e-05,
+      "loss": 2.7874,
+      "step": 204
+    },
+    {
+      "epoch": 0.06932817298388794,
+      "grad_norm": 0.8452678322792053,
+      "learning_rate": 9.990293704951001e-05,
+      "loss": 2.786,
+      "step": 206
+    },
+    {
+      "epoch": 0.07000126204198392,
+      "grad_norm": 0.9482727646827698,
+      "learning_rate": 9.989588262696357e-05,
+      "loss": 2.8156,
+      "step": 208
+    },
+    {
+      "epoch": 0.07067435110007993,
+      "grad_norm": 0.8251766562461853,
+      "learning_rate": 9.988858103057378e-05,
+      "loss": 2.7588,
+      "step": 210
+    },
+    {
+      "epoch": 0.07134744015817593,
+      "grad_norm": 1.211065649986267,
+      "learning_rate": 9.988103229651121e-05,
+      "loss": 2.7623,
+      "step": 212
+    },
+    {
+      "epoch": 0.07202052921627193,
+      "grad_norm": 0.8990377187728882,
+      "learning_rate": 9.987323646217075e-05,
+      "loss": 2.8164,
+      "step": 214
+    },
+    {
+      "epoch": 0.07269361827436793,
+      "grad_norm": 0.9878025054931641,
+      "learning_rate": 9.986519356617132e-05,
+      "loss": 2.7847,
+      "step": 216
+    },
+    {
+      "epoch": 0.07336670733246392,
+      "grad_norm": 0.7551445364952087,
+      "learning_rate": 9.985690364835576e-05,
+      "loss": 2.8111,
+      "step": 218
+    },
+    {
+      "epoch": 0.07403979639055992,
+      "grad_norm": 0.9582260251045227,
+      "learning_rate": 9.984836674979062e-05,
+      "loss": 2.793,
+      "step": 220
+    },
+    {
+      "epoch": 0.07471288544865592,
+      "grad_norm": 0.8087739944458008,
+      "learning_rate": 9.983958291276591e-05,
+      "loss": 2.7464,
+      "step": 222
+    },
+    {
+      "epoch": 0.07538597450675193,
+      "grad_norm": 1.2373522520065308,
+      "learning_rate": 9.983055218079493e-05,
+      "loss": 2.7656,
+      "step": 224
+    },
+    {
+      "epoch": 0.07605906356484793,
+      "grad_norm": 0.9746289849281311,
+      "learning_rate": 9.982127459861408e-05,
+      "loss": 2.7765,
+      "step": 226
+    },
+    {
+      "epoch": 0.07673215262294392,
+      "grad_norm": 0.6946307420730591,
+      "learning_rate": 9.981175021218255e-05,
+      "loss": 2.7491,
+      "step": 228
+    },
+    {
+      "epoch": 0.07740524168103992,
+      "grad_norm": 0.8959107398986816,
+      "learning_rate": 9.980197906868215e-05,
+      "loss": 2.7565,
+      "step": 230
+    },
+    {
+      "epoch": 0.07807833073913592,
+      "grad_norm": 0.9889335036277771,
+      "learning_rate": 9.979196121651716e-05,
+      "loss": 2.7974,
+      "step": 232
+    },
+    {
+      "epoch": 0.07875141979723192,
+      "grad_norm": 0.9802746176719666,
+      "learning_rate": 9.978169670531388e-05,
+      "loss": 2.7772,
+      "step": 234
+    },
+    {
+      "epoch": 0.07942450885532792,
+      "grad_norm": 0.6934760808944702,
+      "learning_rate": 9.977118558592059e-05,
+      "loss": 2.7602,
+      "step": 236
+    },
+    {
+      "epoch": 0.08009759791342393,
+      "grad_norm": 0.8996357917785645,
+      "learning_rate": 9.97604279104072e-05,
+      "loss": 2.7669,
+      "step": 238
+    },
+    {
+      "epoch": 0.08077068697151991,
+      "grad_norm": 0.8844061493873596,
+      "learning_rate": 9.974942373206499e-05,
+      "loss": 2.7458,
+      "step": 240
+    },
+    {
+      "epoch": 0.08144377602961592,
+      "grad_norm": 1.023626685142517,
+      "learning_rate": 9.973817310540638e-05,
+      "loss": 2.7639,
+      "step": 242
+    },
+    {
+      "epoch": 0.08211686508771192,
+      "grad_norm": 0.8241132497787476,
+      "learning_rate": 9.972667608616466e-05,
+      "loss": 2.7457,
+      "step": 244
+    },
+    {
+      "epoch": 0.08278995414580792,
+      "grad_norm": 0.7864794135093689,
+      "learning_rate": 9.971493273129364e-05,
+      "loss": 2.763,
+      "step": 246
+    },
+    {
+      "epoch": 0.08346304320390392,
+      "grad_norm": 1.071751356124878,
+      "learning_rate": 9.970294309896747e-05,
+      "loss": 2.7347,
+      "step": 248
+    },
+    {
+      "epoch": 0.08413613226199991,
+      "grad_norm": 0.8978875279426575,
+      "learning_rate": 9.969070724858031e-05,
+      "loss": 2.7807,
+      "step": 250
+    },
+    {
+      "epoch": 0.08480922132009591,
+      "grad_norm": 0.9984204769134521,
+      "learning_rate": 9.967822524074602e-05,
+      "loss": 2.7399,
+      "step": 252
+    },
+    {
+      "epoch": 0.08548231037819191,
+      "grad_norm": 0.8611739873886108,
+      "learning_rate": 9.966549713729787e-05,
+      "loss": 2.753,
+      "step": 254
+    },
+    {
+      "epoch": 0.08615539943628792,
+      "grad_norm": 0.8647720217704773,
+      "learning_rate": 9.965252300128826e-05,
+      "loss": 2.7224,
+      "step": 256
+    },
+    {
+      "epoch": 0.08682848849438392,
+      "grad_norm": 0.8688477873802185,
+      "learning_rate": 9.963930289698833e-05,
+      "loss": 2.6879,
+      "step": 258
+    },
+    {
+      "epoch": 0.0875015775524799,
+      "grad_norm": 1.1445469856262207,
+      "learning_rate": 9.962583688988778e-05,
+      "loss": 2.739,
+      "step": 260
+    },
+    {
+      "epoch": 0.08817466661057591,
+      "grad_norm": 0.8668599128723145,
+      "learning_rate": 9.961212504669437e-05,
+      "loss": 2.6962,
+      "step": 262
+    },
+    {
+      "epoch": 0.08884775566867191,
+      "grad_norm": 0.905125617980957,
+      "learning_rate": 9.959816743533375e-05,
+      "loss": 2.7239,
+      "step": 264
+    },
+    {
+      "epoch": 0.08952084472676791,
+      "grad_norm": 0.8252028822898865,
+      "learning_rate": 9.958396412494901e-05,
+      "loss": 2.7381,
+      "step": 266
+    },
+    {
+      "epoch": 0.09019393378486391,
+      "grad_norm": 0.7380514740943909,
+      "learning_rate": 9.956951518590043e-05,
+      "loss": 2.7135,
+      "step": 268
+    },
+    {
+      "epoch": 0.09086702284295992,
+      "grad_norm": 0.7395239472389221,
+      "learning_rate": 9.955482068976502e-05,
+      "loss": 2.6954,
+      "step": 270
+    },
+    {
+      "epoch": 0.0915401119010559,
+      "grad_norm": 0.6564229726791382,
+      "learning_rate": 9.953988070933628e-05,
+      "loss": 2.7145,
+      "step": 272
+    },
+    {
+      "epoch": 0.0922132009591519,
+      "grad_norm": 0.7306910157203674,
+      "learning_rate": 9.952469531862378e-05,
+      "loss": 2.6951,
+      "step": 274
+    },
+    {
+      "epoch": 0.09288629001724791,
+      "grad_norm": 0.6810031533241272,
+      "learning_rate": 9.950926459285277e-05,
+      "loss": 2.7201,
+      "step": 276
+    },
+    {
+      "epoch": 0.09355937907534391,
+      "grad_norm": 0.6724168658256531,
+      "learning_rate": 9.949358860846388e-05,
+      "loss": 2.7112,
+      "step": 278
+    },
+    {
+      "epoch": 0.09423246813343991,
+      "grad_norm": 0.7065703272819519,
+      "learning_rate": 9.947766744311268e-05,
+      "loss": 2.6884,
+      "step": 280
+    },
+    {
+      "epoch": 0.0949055571915359,
+      "grad_norm": 0.8231908679008484,
+      "learning_rate": 9.946150117566931e-05,
+      "loss": 2.7286,
+      "step": 282
+    },
+    {
+      "epoch": 0.0955786462496319,
+      "grad_norm": 0.9570270776748657,
+      "learning_rate": 9.944508988621812e-05,
+      "loss": 2.7166,
+      "step": 284
+    },
+    {
+      "epoch": 0.0962517353077279,
+      "grad_norm": 0.9357023239135742,
+      "learning_rate": 9.94284336560572e-05,
+      "loss": 2.6768,
+      "step": 286
+    },
+    {
+      "epoch": 0.0969248243658239,
+      "grad_norm": 0.6350796222686768,
+      "learning_rate": 9.941153256769809e-05,
+      "loss": 2.6921,
+      "step": 288
+    },
+    {
+      "epoch": 0.09759791342391991,
+      "grad_norm": 0.6700872778892517,
+      "learning_rate": 9.939438670486525e-05,
+      "loss": 2.6847,
+      "step": 290
+    },
+    {
+      "epoch": 0.0982710024820159,
+      "grad_norm": 0.6851752400398254,
+      "learning_rate": 9.937699615249572e-05,
+      "loss": 2.6586,
+      "step": 292
+    },
+    {
+      "epoch": 0.0989440915401119,
+      "grad_norm": 0.7098946571350098,
+      "learning_rate": 9.935936099673871e-05,
+      "loss": 2.6793,
+      "step": 294
+    },
+    {
+      "epoch": 0.0996171805982079,
+      "grad_norm": 0.680543839931488,
+      "learning_rate": 9.934148132495511e-05,
+      "loss": 2.6763,
+      "step": 296
+    },
+    {
+      "epoch": 0.1002902696563039,
+      "grad_norm": 0.6832155585289001,
+      "learning_rate": 9.932335722571709e-05,
+      "loss": 2.6768,
+      "step": 298
+    },
+    {
+      "epoch": 0.1009633587143999,
+      "grad_norm": 0.8236553072929382,
+      "learning_rate": 9.930498878880768e-05,
+      "loss": 2.6738,
+      "step": 300
+    }
+  ],
+  "logging_steps": 2,
+  "max_steps": 2972,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 300,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 3.377550336196608e+17,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}

q2.5-eu-en/checkpoint-300/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

q2.5/checkpoint-12000/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

q2.5/checkpoint-12000/chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

q2.5/checkpoint-12000/config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 896,
+  "initializer_range": 0.02,
+  "intermediate_size": 4864,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 24,
+  "model_type": "qwen2",
+  "num_attention_heads": 14,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": true,
+  "use_mrope": false,
+  "use_sliding_window": false,
+  "vocab_size": 151665
+}

q2.5/checkpoint-12000/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "transformers_version": "4.52.4"
+}

q2.5/checkpoint-12000/latest

@@ -0,0 +1 @@
+global_step12000