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 3, 21:30 UTC
[Bugfix] Fix double reduce in flashinfer_nvlink_two_sided and flashinfer_nvlink_one_sided backends (#41382)
Signed-off-by: amitz-nv <203509407+amitz-nv@users.noreply.github.com>
amitz-nv committed
May 12, 2026 at 10:47 UTC
ef34592a1ac7a203a490b1e067d29dfaa6a0b2a9
2 files changed
+2
-2
vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py
+1
-1
@@ -67,7 +67,7 @@ class FlashInferNVLinkOneSidedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeMo
67
return self.num_dispatchers_
68
69
def output_is_reduced(self) -> bool:
70
- return False
70
+ return True
71
72
def topk_indices_dtype(self) -> torch.dtype | None:
73
return torch.int32
vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py
+1
-1
@@ -50,7 +50,7 @@ class FlashInferNVLinkTwoSidedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeMo
50
return self.num_dispatchers_
51
52
def output_is_reduced(self) -> bool:
53
- return False
53
+ return True
54
55
def _apply_router_weight_on_input(
56
self,