From ef34592a1ac7a203a490b1e067d29dfaa6a0b2a9 Mon Sep 17 00:00:00 2001 From: amitz-nv <203509407+amitz-nv@users.noreply.github.com> Date: Tue, 12 May 2026 10:47:47 +0300 Subject: [PATCH] [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> --- .../fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py | 2 +- .../fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py b/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py index 254153a27c..e49d8b2624 100644 --- a/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py +++ b/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py @@ -67,7 +67,7 @@ class FlashInferNVLinkOneSidedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeMo return self.num_dispatchers_ def output_is_reduced(self) -> bool: - return False + return True def topk_indices_dtype(self) -> torch.dtype | None: return torch.int32 diff --git a/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py b/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py index b975e6cc5e..f89f375709 100644 --- a/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py +++ b/vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py @@ -50,7 +50,7 @@ class FlashInferNVLinkTwoSidedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeMo return self.num_dispatchers_ def output_is_reduced(self) -> bool: - return False + return True def _apply_router_weight_on_input( self,