From b0d1c21d03f3e921f84bbcf4e111df8ce976a4bc Mon Sep 17 00:00:00 2001 From: Lee Nau Date: Thu, 6 Nov 2025 11:53:58 -0800 Subject: [PATCH] [fix] Only enable flashinfer all reduce fusion by default for single-node servers (#12724) --- python/sglang/srt/server_args.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index 6e3ec921b..d322a1d9c 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -907,7 +907,8 @@ class ServerArgs: logger.info( "Use trtllm_mla as attention backend on sm100 for DeepseekV3ForCausalLM" ) - if not self.enable_dp_attention: + # workaround for https://github.com/flashinfer-ai/flashinfer/issues/2006 + if not self.enable_dp_attention and self.nnodes == 1: self.enable_flashinfer_allreduce_fusion = True logger.info( "Enable FlashInfer AllReduce Fusion on sm100 for DeepseekV3ForCausalLM" @@ -964,7 +965,8 @@ class ServerArgs: ) if is_blackwell_supported(): - if not self.enable_dp_attention: + # workaround for https://github.com/flashinfer-ai/flashinfer/issues/2006 + if not self.enable_dp_attention and self.nnodes == 1: self.enable_flashinfer_allreduce_fusion = True logger.info( "Enable FlashInfer AllReduce Fusion on sm100 for GptOssForCausalLM"