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

Fix kernel selection in biased_grouped_topk_gpu (#17325)

Browse Source
This commit is contained in:
yudian0504
2026-01-19 14:06:11 +08:00
committed by GitHub
parent 858a4d659b
commit 9fe56cd0fb
1 changed files with 0 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/moe/topk.py
-1
View File
@@ -795,7 +795,6 @@ def biased_grouped_topk_gpu(
elif ( elif (
_is_cuda _is_cuda
and num_fused_shared_experts == 0
# moe_fused_gate kernel ensures that num_experts/num_expert_group does not exceed MAX_VPT=32 now. And when kernel can handle MAX_VPT > 32, we can remove this assertion. # moe_fused_gate kernel ensures that num_experts/num_expert_group does not exceed MAX_VPT=32 now. And when kernel can handle MAX_VPT > 32, we can remove this assertion.
and experts_per_group <= 32 and experts_per_group <= 32
and is_power_of_two(num_experts) and is_power_of_two(num_experts)