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

[Bugfix] Fix top-1 MoE routing with non-unit scaling (#40187)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2026-09-18 22:53:13 +08:00
committed by GitHub
parent 81363bf8cb
commit 7714b182f2
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe.py
+5 -2
View File
@@ -813,8 +813,11 @@ def _fused_moe_kernel_sequence(
) )
# LoRA hooks force the second kernel to write to intermediate_cache3 so # LoRA hooks force the second kernel to write to intermediate_cache3 so
# hooks.after_down can inspect/modify it before reduction. # hooks.after_down can inspect/modify it before reduction. Non-unit routed
_use_intermediate = not no_combine and (topk != 1 or hooks) # scaling also needs the intermediate because the reduction applies it.
_use_intermediate = not no_combine and (
topk != 1 or hooks or routed_scaling_factor not in (None, 1.0)
)
out_slice = None out_slice = None
if use_fused_moe_sum_all_reduce: if use_fused_moe_sum_all_reduce: