xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-04 11:53:06 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bug] Fix DeepSeek-V2.5-1210-FP8 issue (#27267)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye 2025-10-22 11:00:10 -04:00 committed by GitHub
parent 4ca13a8667
commit 1c160841ea
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
vllm/model_executor/layers/fused_moe/cutlass_moe.py
View File

@ -511,13 +511,19 @@ def cutlass_moe_fp8(
assert quant_config is not None
if quant_config.a1_scale is not None:
assert quant_config.per_act_token_quant == quant_config.a1_scale.numel() != 1
assert quant_config.per_act_token_quant == (quant_config.a1_scale.numel() != 1)
if quant_config.a2_scale is not None:
assert quant_config.per_act_token_quant == quant_config.a2_scale.numel() != 1
assert quant_config.per_act_token_quant == (quant_config.a2_scale.numel() != 1)
assert quant_config.w1_scale is None or (
quant_config.per_out_ch_quant == (quant_config.w1_scale.size(1) == w1_q.size(1))
)
if quant_config.w1_scale is not None:
if quant_config.per_out_ch_quant:
assert quant_config.w1_scale.dim() > 1 and quant_config.w1_scale.size(
1
) == w1_q.size(1)
else:
assert (
quant_config.w1_scale.dim() == 1 or quant_config.w1_scale.size(1) == 1
)
num_experts = global_num_experts if global_num_experts != -1 else w1_q.size(0)