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 13, 04:30 UTC
[BugFix] Fix misprint introduced by modular_kernel refactoring. (#28728)
Signed-off-by: Andrey Khalyavin <halyavin@yandex-team.ru>
Andrey Khalyavin committed
Nov 14, 2025 at 21:58 UTC
fd4555089a7ea3094499d9a6a9cec1c1b6903674
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/modular_kernel.py
+1
-1
@@ -1060,7 +1060,7 @@ class FusedMoEModularKernel(torch.nn.Module):
1060
global_num_experts=global_num_experts,
1061
expert_map=expert_map,
1062
a1q_scale=_slice_scales(a1q_scale, s, e),
1063
- a2_scale=_slice_scales(self.fused_experts.a2_scale, e, e),
1063
+ a2_scale=_slice_scales(self.fused_experts.a2_scale, s, e),
1064
workspace13=workspace13,
1065
workspace2=workspace2,
1066
expert_tokens_meta=c_expert_tokens_meta,