model:
  name: Qwen3-VL-32B-Instruct
  model_type: qwen3_vl

paths:
  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-32B-Instruct
  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-32B-Instruct-w4a8

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+.self_attn.v_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.o_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
    NN: ~
  batch_size: 64
  block_size: 128
  group_size: -1
  symmetric: True
  class:
    int4: ~
    int8:
      - "re:model.language_model.layers.\\d+.self_attn.v_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.o_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
  ignore: 
    - "re:.*visual.*"
    - "re:.*norm.*"
    - "re:lm_head.*"
    - "re:.*embed_tokens.*"

dequant:
  ignore:
    - "re:lm_head.*"
    - "re:model.embed_tokens.*"
    - "re:model.layers.\\d+.input_layernorm.*"
    - "re:model.layers.\\d+.mlp.gate.*"
    - "re:model.layers.\\d+.post_attention_layernorm.*"
    - "re:model.layers.\\d+.self_attn.k_norm.*"
    - "re:model.layers.\\d+.self_attn.q_norm.*"
    - "re:model.norm.*"