model:
  name: Qwen3.5-35B-A3B
  model_type: qwen3_5_moe

paths:
  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-35B-A3B
  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-35B-A3B-w4a8

runtime:
  num_workers: 16
  mode: quant # dequant | dequant+quant | quant (only support bf16)

retry_files: ~

quant:
  version: 2
  format: 
    TN:
      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.k_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: 
      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
    int8:
      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
  ignore: 
    - "re:lm_head.*"
    - "re:.*embed_tokens.*"
    - "re:.*visual.*"
    - "re:.*mtp.*"
    - "re:.*norm.*"
    - "re:.*linear_attn.*"
    - "re:model.language_model.layers.\\d+.mlp.gate\\..*"
    - "re:.*shared_expert.*"

dequant:
  ignore: ~