Fix kernel selection in biased_grouped_topk_gpu (#17325)

2026-01-19 14:06:11 +08:00
parent 858a4d659b
commit 9fe56cd0fb
1 changed files with 0 additions and 1 deletions
--- a/python/sglang/srt/layers/moe/topk.py
+++ b/python/sglang/srt/layers/moe/topk.py
@@ -795,7 +795,6 @@ def biased_grouped_topk_gpu(

    elif (
        _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.
        and experts_per_group <= 32
        and is_power_of_two(num_experts)