From 74a166cb863c6afea9e5f83772b09badf7486d11 Mon Sep 17 00:00:00 2001 From: b8zhong Date: Fri, 2 Jan 2026 19:43:17 -0800 Subject: [PATCH] [Fix] Only add SM90 and SM100 to check for auto-enabling TRT Allreduce Fusion (#16283) --- python/sglang/srt/server_args.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index 43bbea6c3..c94b47238 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -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