A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bug] Fix R1 Accuracy 0 Bug (#23294)
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>
Wentao Ye committed
Aug 21, 2025 at 13:11 UTC
48bfb0c9b733aab69f4c7d77051e20a80c8d088f
1 file changed
+15
-7
vllm/model_executor/layers/quantization/fp8.py
+15
-7
@@ -1099,8 +1099,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
1099
apply_router_weight_on_input=apply_router_weight_on_input,
1100
)
1101
else:
1102
- from vllm.model_executor.layers.fused_moe import fused_experts
1103
- return fused_experts(
1102
+ common_kwargs = dict(
1103
hidden_states=x,
1104
w1=layer.w13_weight,
1105
w2=layer.w2_weight,
@@ -1117,11 +1116,20 @@ class Fp8MoEMethod(FusedMoEMethodBase):
1116
if self.block_quant else layer.w2_weight_scale),
1117
a1_scale=layer.w13_input_scale,
1118
a2_scale=layer.w2_input_scale,
1120
- use_fp8_w8a8=True,
1121
- block_shape=self.quant_config.weight_block_size,
1122
- allow_deep_gemm=self.allow_deep_gemm,
1123
- allow_cutlass_block_scaled_grouped_gemm=(
1124
- self.allow_cutlass_block_scaled_grouped_gemm))
1119
+ )
1120
+
1121
+ if self.fused_experts is not None:
1122
+ return self.fused_experts(**common_kwargs)
1123
+ else:
1124
+ from vllm.model_executor.layers.fused_moe import fused_experts
1125
+ return fused_experts(
1126
+ **common_kwargs,
1127
+ use_fp8_w8a8=True,
1128
+ block_shape=self.quant_config.weight_block_size,
1129
+ allow_deep_gemm=self.allow_deep_gemm,
1130
+ allow_cutlass_block_scaled_grouped_gemm=(
1131
+ self.allow_cutlass_block_scaled_grouped_gemm),
1132
+ )
1133
1134
1135
class Fp8KVCacheMethod(BaseKVCacheMethod):