52 lines
1.1 KiB
YAML
52 lines
1.1 KiB
YAML
|
model:
|
||
|
|
name: Qwen3-Next-80B-A3B-Instruct
|
||
|
|
model_type: qwen3_moe
|
||
|
|
|
||
|
|
paths:
|
||
|
|
checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Next-80B-A3B-Instruct
|
||
|
|
output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Next-80B-A3B-Instruct-w8a8
|
||
|
|
|
||
|
|
runtime:
|
||
|
|
num_workers: 16
|
||
|
|
mode: dequant # dequant | dequant+quant | quant (only support bf16)
|
||
|
|
|
||
|
|
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:.*linear_attn.*"
|
||
|
|
- "re:.*self_attn.*"
|
||
|
|
- "re:.*shared_expert.*"
|
||
|
|
- "re:.*linear_attn.in_proj_ba.*"
|
||
|
|
- "re:.*conv1d.*"
|
||
|
|
- "re:.*A_log.*"
|
||
|
|
- "re:.*shared_expert_gate.*"
|
||
|
|
- "re:.*bias.*"
|
||
|
|
- "re:.*norm.*"
|
||
|
|
- "re:.*mlp.gate.*"
|
||
|
|
- "re:.*embed_tokens.*"
|
||
|
|
- "re:.*lm_head.*"
|
||
|
|
- "re:.*mtp.*"
|
||
|
|
|
||
|
|
dequant:
|
||
|
|
ignore:
|
||
|
|
- "re:.*linear_attn.in_proj_ba*"
|
||
|
|
- "re:.*conv1d.*"
|
||
|
|
- "re:.*A_log.*"
|
||
|
|
- "re:.*shared_expert_gate.*"
|
||
|
|
- "re:.*bias.*"
|
||
|
|
- "re:.*norm.*"
|
||
|
|
- "re:.*mlp.gate.*"
|
||
|
|
- "re:.*embed_tokens.*"
|
||
|
|
- "re:.*lm_head.*"
|