minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[hotfix] Reenable all reduce fusion on sm100 (#17591)

Browse Source
This commit is contained in:
Baizhou Zhang
2026-01-22 23:36:38 +08:00
committed by GitHub
parent f7a0bcda1e
commit 283a2daeaa
1 changed files with 2 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/communicator.py
+2 -3
View File
@@ -88,10 +88,9 @@ FUSE_ALLREDUCE_MAX_BATCH_SIZE = 2048
def apply_flashinfer_allreduce_fusion(batch_size: int): def apply_flashinfer_allreduce_fusion(batch_size: int):
return ( return (
# TODO: flashinfer 0.6.1 caused performance regression on sm100 for allreduce fusion # NOTE: 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.
# Ref: https://github.com/sgl-project/sglang/issues/17237 # Ref: https://github.com/sgl-project/sglang/issues/17237
_is_sm90_supported (_is_sm90_supported or _is_sm100_supported)
and _is_flashinfer_available and _is_flashinfer_available
and batch_size > 0 and batch_size > 0
and batch_size <= FUSE_ALLREDUCE_MAX_BATCH_SIZE and batch_size <= FUSE_ALLREDUCE_MAX_BATCH_SIZE