xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-23 20:07:15 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Correct per_act_token in CompressedTensorsW8A8Fp8MoECutlassM… (#20937)

Browse Source 
Signed-off-by: Ming Yang <minos.future@gmail.com>
This commit is contained in:
Ming Yang 2025-07-15 19:53:42 -07:00 committed by GitHub
parent 3ed94f9d0a
commit fcb9f879c1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
View File

@ -929,10 +929,8 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
scoring_func=scoring_func, scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias) e_score_correction_bias=e_score_correction_bias)
a1_scale = layer.w13_input_scale per_act_token = (
a2_scale = layer.w2_input_scale self.input_quant.strategy == QuantizationStrategy.TOKEN)
per_act_token = a1_scale.numel() != 1 if a1_scale is not None else (
a2_scale.numel() != 1 if a2_scale is not None else False)
if self.fused_experts is None: if self.fused_experts is None:
# If no modular kernel is provided, use cutlass_moe_fp8 # If no modular kernel is provided, use cutlass_moe_fp8
@ -950,8 +948,8 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
expert_map=None if self.disable_expert_map else expert_map, expert_map=None if self.disable_expert_map else expert_map,
w1_scale=layer.w13_weight_scale, w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale, w2_scale=layer.w2_weight_scale,
a1_scale=a1_scale, a1_scale=layer.w13_input_scale,
a2_scale=a2_scale, a2_scale=layer.w2_input_scale,
) )
else: else:
return self.fused_experts( return self.fused_experts(