minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Fix Qwen3 MoE: also guard EP all-reduce with not use_reduce_scatter (follow-up to #23731) (#23734)

Browse Source 
Co-authored-by: Byron Hsu <byron@periodiclabs.ai>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Byron Hsu
2026-04-25 20:35:52 -07:00
committed by GitHub
co-authored by Byron Hsu Claude Opus 4.7
parent 714173555c
commit 71029abd64
1 changed files with 1 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_moe.py
+1
View File
@@ -335,6 +335,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
if ( if (
self.ep_size > 1 self.ep_size > 1
and not should_allreduce_fusion and not should_allreduce_fusion
and not use_reduce_scatter
and not should_use_dp_reduce_scatterv() and not should_use_dp_reduce_scatterv()
): ):
final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states) final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)