From 71029abd640f9b90dbf2177119a04c14e997c3dd Mon Sep 17 00:00:00 2001 From: Byron Hsu Date: Sat, 25 Apr 2026 20:35:52 -0700 Subject: [PATCH] Fix Qwen3 MoE: also guard EP all-reduce with not use_reduce_scatter (follow-up to #23731) (#23734) Co-authored-by: Byron Hsu Co-authored-by: Claude Opus 4.7 (1M context) --- python/sglang/srt/models/qwen3_moe.py | 1 + 1 file changed, 1 insertion(+) diff --git a/python/sglang/srt/models/qwen3_moe.py b/python/sglang/srt/models/qwen3_moe.py index 21412a161..dd9598c36 100644 --- a/python/sglang/srt/models/qwen3_moe.py +++ b/python/sglang/srt/models/qwen3_moe.py @@ -335,6 +335,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module): if ( self.ep_size > 1 and not should_allreduce_fusion + and not use_reduce_scatter and not should_use_dp_reduce_scatterv() ): final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)