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

[B300] Enable FlashInfer allreduce for Qwen3-VL MoE (#28786)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2026-06-22 22:38:45 +08:00
committed by GitHub
parent cee1caaf47
commit b43bd6824f
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+6 -4
View File
@@ -3131,10 +3131,11 @@ class ServerArgs:
# Auto-enable FlashInfer AllReduce Fusion on SM100 only, for models with # Auto-enable FlashInfer AllReduce Fusion on SM100 only, for models with
# explicit support (DeepseekV3, GptOss, Glm4Moe, MistralLarge3, # explicit support (DeepseekV3, GptOss, Glm4Moe, MistralLarge3,
# Qwen3/Qwen3Next/Qwen3.5 MoE families). SM90 is not auto-enabled because # Qwen3/Qwen3-VL/Qwen3Next/Qwen3.5 MoE families). SM90 is not
# auto resolves to mnnvl, which requires a working NVLink multicast fabric # auto-enabled because auto resolves to mnnvl, which requires a working
# that SM90 nodes do not reliably have; SM90 users can opt in explicitly # NVLink multicast fabric that SM90 nodes do not reliably have; SM90
# via --flashinfer-allreduce-fusion-backend. # users can opt in explicitly via
# --flashinfer-allreduce-fusion-backend.
if ( if (
self.flashinfer_allreduce_fusion_backend is None self.flashinfer_allreduce_fusion_backend is None
and model_arch and model_arch
@@ -3147,6 +3148,7 @@ class ServerArgs:
"Glm4MoeLiteForCausalLM", "Glm4MoeLiteForCausalLM",
"MistralLarge3ForCausalLM", "MistralLarge3ForCausalLM",
"Qwen3MoeForCausalLM", "Qwen3MoeForCausalLM",
"Qwen3VLMoeForConditionalGeneration",
"Qwen3NextForCausalLM", "Qwen3NextForCausalLM",
"KimiK25ForConditionalGeneration", "KimiK25ForConditionalGeneration",
"Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForConditionalGeneration",