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

[fix] Only enable flashinfer all reduce fusion by default for single-node servers (#12724)

Browse Source
This commit is contained in:
Lee Nau
2025-11-06 11:53:58 -08:00
committed by GitHub
parent b07c5e4080
commit b0d1c21d03
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/server_args.py
View File

@@ -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"