From 283a2daeaa88532299fc5a9c6db5bcab41f6ef97 Mon Sep 17 00:00:00 2001 From: Baizhou Zhang Date: Thu, 22 Jan 2026 23:36:38 +0800 Subject: [PATCH] [hotfix] Reenable all reduce fusion on sm100 (#17591) --- python/sglang/srt/layers/communicator.py | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/python/sglang/srt/layers/communicator.py b/python/sglang/srt/layers/communicator.py index 0115a189c..ebd0d3bea 100644 --- a/python/sglang/srt/layers/communicator.py +++ b/python/sglang/srt/layers/communicator.py @@ -88,10 +88,9 @@ FUSE_ALLREDUCE_MAX_BATCH_SIZE = 2048 def apply_flashinfer_allreduce_fusion(batch_size: int): return ( - # TODO: flashinfer 0.6.1 caused performance regression on sm100 for allreduce fusion - # Temporarily disable it on sm100. Add it back after its performance is restored. + # NOTE: flashinfer 0.6.1 caused performance regression on sm100 for allreduce fusion # Ref: https://github.com/sgl-project/sglang/issues/17237 - _is_sm90_supported + (_is_sm90_supported or _is_sm100_supported) and _is_flashinfer_available and batch_size > 0 and batch_size <= FUSE_ALLREDUCE_MAX_BATCH_SIZE