From a54d75bf2e954564a6c9cee56405e813d78a6d1d Mon Sep 17 00:00:00 2001 From: Baizhou Zhang Date: Wed, 21 Jan 2026 13:54:09 +0800 Subject: [PATCH] [Fix] Set fa3 as default MHA backend on Hopper (#17425) --- python/sglang/srt/server_args.py | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index cef7baede..837d0de2e 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -45,7 +45,6 @@ from sglang.srt.utils.common import ( get_quantization_config, is_blackwell_supported, is_cuda, - is_fa3_default_architecture, is_flashinfer_available, is_hip, is_hopper_with_cuda_12_3, @@ -1630,11 +1629,10 @@ class ServerArgs: if not use_mla_backend: # MHA architecture - if ( - is_hopper_with_cuda_12_3() - and is_no_spec_infer_or_topk_one(self) - and is_fa3_default_architecture(self.model_config.hf_config) - ): + if is_hopper_with_cuda_12_3() and is_no_spec_infer_or_topk_one(self): + # Note: flashinfer 0.6.1 caused performance regression on Hopper attention kernel + # Before the kernel is fixed, we choose fa3 as the default backend on Hopper MHA + # ref: https://github.com/sgl-project/sglang/issues/17411 self.attention_backend = "fa3" elif ( is_sm100_supported()