model:
  name: Qwen3.5-27B
  model_type: qwen3_5

paths:
  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-27B
  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-27B-w8a8

runtime:
  num_workers: 8
  mode: quant # dequant | dequant+quant | quant (only support bf16)

retry_files: ~

quant:
  version: 2
  format:
    TN:
      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
    NN: ~
  batch_size: 64
  block_size: 128
  group_size: -1
  symmetric: True
  class:
    int4: ~
    int8:
      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
  ignore: 
    - "re:lm_head.*"
    - "re:.*visual.*"
    - "re:.*mtp.*"
    - "re:.*linear_attn.*"
    - "re:.*norm.*"
    - "re:.*embed_tokens.*"

dequant:
  ignore: ~