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

[Fix] Set fa3 as default MHA backend on Hopper (#17425)

Browse Source
This commit is contained in:
Baizhou Zhang
2026-01-21 13:54:09 +08:00
committed by GitHub
parent 3321eb4efa
commit a54d75bf2e
1 changed files with 4 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+4 -6
View File
@@ -45,7 +45,6 @@ from sglang.srt.utils.common import (
get_quantization_config, get_quantization_config,
is_blackwell_supported, is_blackwell_supported,
is_cuda, is_cuda,
is_fa3_default_architecture,
is_flashinfer_available, is_flashinfer_available,
is_hip, is_hip,
is_hopper_with_cuda_12_3, is_hopper_with_cuda_12_3,
@@ -1630,11 +1629,10 @@ class ServerArgs:
if not use_mla_backend: if not use_mla_backend:
# MHA architecture # MHA architecture
if ( if is_hopper_with_cuda_12_3() and is_no_spec_infer_or_topk_one(self):
is_hopper_with_cuda_12_3() # Note: flashinfer 0.6.1 caused performance regression on Hopper attention kernel
and is_no_spec_infer_or_topk_one(self) # Before the kernel is fixed, we choose fa3 as the default backend on Hopper MHA
and is_fa3_default_architecture(self.model_config.hf_config) # ref: https://github.com/sgl-project/sglang/issues/17411
):
self.attention_backend = "fa3" self.attention_backend = "fa3"
elif ( elif (
is_sm100_supported() is_sm100_supported()