config.json

{
    "dfloat11_config": {
      "bytes_per_thread": 8,
      "pattern_dict": {
        "double_stream_blocks\\.\\d+": [
            "block.adaLN_modulation.1",
            "block.attn1.to_q",
            "block.attn1.to_k",
            "block.attn1.to_v",
            "block.attn1.to_out",
            "block.attn1.to_q_t",
            "block.attn1.to_k_t",
            "block.attn1.to_v_t",
            "block.attn1.to_out_t",
            "block.ff_i.shared_experts.w1",
            "block.ff_i.shared_experts.w2",
            "block.ff_i.shared_experts.w3",
            "block.ff_i.experts.0.w1",
            "block.ff_i.experts.0.w2",
            "block.ff_i.experts.0.w3",
            "block.ff_i.experts.1.w1",
            "block.ff_i.experts.1.w2",
            "block.ff_i.experts.1.w3",
            "block.ff_i.experts.2.w1",
            "block.ff_i.experts.2.w2",
            "block.ff_i.experts.2.w3",
            "block.ff_i.experts.3.w1",
            "block.ff_i.experts.3.w2",
            "block.ff_i.experts.3.w3",
            "block.ff_t.w1",
            "block.ff_t.w2",
            "block.ff_t.w3"
          ],
        "single_stream_blocks\\.\\d+": [
            "block.adaLN_modulation.1",
            "block.attn1.to_q",
            "block.attn1.to_k",
            "block.attn1.to_v",
            "block.attn1.to_out",
            "block.ff_i.shared_experts.w1",
            "block.ff_i.shared_experts.w2",
            "block.ff_i.shared_experts.w3",
            "block.ff_i.experts.0.w1",
            "block.ff_i.experts.0.w2",
            "block.ff_i.experts.0.w3",
            "block.ff_i.experts.1.w1",
            "block.ff_i.experts.1.w2",
            "block.ff_i.experts.1.w3",
            "block.ff_i.experts.2.w1",
            "block.ff_i.experts.2.w2",
            "block.ff_i.experts.2.w3",
            "block.ff_i.experts.3.w1",
            "block.ff_i.experts.3.w2",
            "block.ff_i.experts.3.w3"
          ]
      },
      "threads_per_block": [
        512
      ],
      "version": "0.2.0"
    },
    "model_type": "llama"
  }