chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Fix] Only add SM90 and SM100 to check for auto-enabling TRT Allreduce Fusion (#16283)

Browse Source
This commit is contained in:
b8zhong
2026-01-02 19:43:17 -08:00
committed by GitHub
parent 888e126ac9
commit 74a166cb86
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/server_args.py
View File

@@ -1516,7 +1516,7 @@ class ServerArgs:
f"Overlap scheduler is disabled when using sparse head for embedding model."
)
# TRTLLM AllReduce Fusion supports SM90/100/120, enable it by default
# TRTLLM AllReduce Fusion supports SM90/100, enable it by default
# for models with explicit support (DeepseekV3, GptOss, Glm4Moe, Qwen3Moe)
# TODO: currently, it is only supported in the single node scenario. https://github.com/flashinfer-ai/flashinfer/issues/2006
# TODO: there is currently a bug on H20 device specifically, https://github.com/flashinfer-ai/flashinfer/issues/2204
@@ -1533,7 +1533,7 @@ class ServerArgs:
"Glm4MoeForCausalLM",
"Qwen3MoeForCausalLM",
]
and (is_sm90_supported() or is_blackwell_supported())
and (is_sm90_supported() or is_sm100_supported())
and not self.enable_dp_attention
and self.nnodes == 1
and not is_h20_device