52 lines
1.2 KiB
YAML
52 lines
1.2 KiB
YAML
|
model:
|
||
|
|
name: Qwen3-30B-A3B-Instruct-2507
|
||
|
|
model_type: qwen3_vl_moe
|
||
|
|
|
||
|
|
paths:
|
||
|
|
checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-30B-A3B-Instruct
|
||
|
|
output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-30B-A3B-Instruct-w4a8
|
||
|
|
|
||
|
|
runtime:
|
||
|
|
num_workers: 16
|
||
|
|
mode: quant # dequant | dequant+quant | quant (only support bf16)
|
||
|
|
|
||
|
|
retry_files: ~
|
||
|
|
|
||
|
|
quant:
|
||
|
|
version: 2
|
||
|
|
format:
|
||
|
|
TN:
|
||
|
|
- "re:.*experts.*"
|
||
|
|
NN: ~
|
||
|
|
batch_size: 64
|
||
|
|
block_size: 128
|
||
|
|
group_size: -1
|
||
|
|
symmetric: True
|
||
|
|
class:
|
||
|
|
int4: ~
|
||
|
|
int8:
|
||
|
|
- "re:.*experts.*"
|
||
|
|
ignore:
|
||
|
|
- "re:lm_head.*"
|
||
|
|
- "re:.*embed_tokens.*"
|
||
|
|
- "re:.*input_layernorm.*"
|
||
|
|
- "re:.*post_attention_layernorm.*"
|
||
|
|
- "re:.*k_norm.*"
|
||
|
|
- "re:.*q_norm.*"
|
||
|
|
- "re:.*mlp\\.gate.*"
|
||
|
|
- "re:.*\\.norm.*"
|
||
|
|
- "re:.*visual.*"
|
||
|
|
- "re:.*k_proj.*"
|
||
|
|
- "re:.*o_proj.*"
|
||
|
|
- "re:.*v_proj.*"
|
||
|
|
- "re:.*q_proj.*"
|
||
|
|
|
||
|
|
dequant:
|
||
|
|
ignore:
|
||
|
|
- "model.language_model.embed_tokens.weight"
|
||
|
|
- "lm_head.weight"
|
||
|
|
- "re:model.language_model.layers.\\d+.input_layernorm.weight"
|
||
|
|
- "re:model.language_model.layers.\\d+.post_attention_layernorm.weight"
|
||
|
|
- "re:model.language_model.norm.weight"
|
||
|
|
- "re:model.visual.*"
|