xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-20 15:17:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bug] Fix R1 Accuracy 0 Bug (#23294)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Wentao Ye 2025-08-21 13:11:28 -04:00 committed by GitHub
parent f8ce022948
commit 48bfb0c9b7
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 15 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
vllm/model_executor/layers/quantization/fp8.py
View File

@ -1099,8 +1099,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
apply_router_weight_on_input=apply_router_weight_on_input, apply_router_weight_on_input=apply_router_weight_on_input,
) )
else: else:
from vllm.model_executor.layers.fused_moe import fused_experts common_kwargs = dict(
return fused_experts(
hidden_states=x, hidden_states=x,
w1=layer.w13_weight, w1=layer.w13_weight,
w2=layer.w2_weight, w2=layer.w2_weight,
@ -1117,11 +1116,20 @@ class Fp8MoEMethod(FusedMoEMethodBase):
if self.block_quant else layer.w2_weight_scale), if self.block_quant else layer.w2_weight_scale),
a1_scale=layer.w13_input_scale, a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale, a2_scale=layer.w2_input_scale,
use_fp8_w8a8=True, )
block_shape=self.quant_config.weight_block_size,
allow_deep_gemm=self.allow_deep_gemm, if self.fused_experts is not None:
allow_cutlass_block_scaled_grouped_gemm=( return self.fused_experts(**common_kwargs)
self.allow_cutlass_block_scaled_grouped_gemm)) else:
from vllm.model_executor.layers.fused_moe import fused_experts
return fused_experts(
**common_kwargs,
use_fp8_w8a8=True,
block_shape=self.quant_config.weight_block_size,
allow_deep_gemm=self.allow_deep_gemm,
allow_cutlass_block_scaled_grouped_gemm=(
self.allow_cutlass_block_scaled_grouped_gemm),
)
class Fp8KVCacheMethod(BaseKVCacheMethod): class Fp8KVCacheMethod(BaseKVCacheMethod):