58 lines
1.8 KiB
YAML
58 lines
1.8 KiB
YAML
model:
|
|
name: Qwen3-Coder-30B-A3B-Instruct-FP8
|
|
model_type: qwen3_moe
|
|
|
|
paths:
|
|
checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Omni-30B-A3B-Thinking
|
|
output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Omni-30B-A3B-Thinking-w4a8
|
|
|
|
runtime:
|
|
num_workers: 8
|
|
mode: quant # dequant | dequant+quant | quant (only support bf16)
|
|
|
|
retry_files: ~
|
|
|
|
quant:
|
|
version: 2
|
|
format:
|
|
TN:
|
|
- "re:.*model.layers.\\d+.mlp.experts.\\d+.down_proj"
|
|
- "re:.*model.layers.\\d+.mlp.experts.\\d+.gate_proj"
|
|
- "re:.*model.layers.\\d+.mlp.experts.\\d+.up_proj"
|
|
NN: ~
|
|
batch_size: 64
|
|
block_size: 128
|
|
group_size: -1
|
|
symmetric: True
|
|
class:
|
|
int4:
|
|
- "re:.*model.layers.\\d+.mlp.experts.\\d+.down_proj"
|
|
- "re:.*model.layers.\\d+.mlp.experts.\\d+.gate_proj"
|
|
- "re:.*model.layers.\\d+.mlp.experts.\\d+.up_proj"
|
|
int8: ~
|
|
ignore:
|
|
- "re:.*lm_head.*"
|
|
- "re:.*model.norm.*"
|
|
- "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.k_proj.*"
|
|
- "re:.*model.layers.\\d+.self_attn.o_proj.*"
|
|
- "re:.*model.layers.\\d+.self_attn.q_norm.*"
|
|
- "re:.*model.layers.\\d+.self_attn.q_proj.*"
|
|
- "re:.*model.layers.\\d+.self_attn.v_proj.*"
|
|
- "re:.*audio_tower.*"
|
|
- "re:.*visual.*"
|
|
|
|
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.*" |