From bed05878f6c8f3a7e730ddef450b728cbe5ad308 Mon Sep 17 00:00:00 2001
From: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>
Date: Fri, 18 Apr 2025 15:17:34 +0800
Subject: [PATCH] fix kimi vl running bug after rebase main (#5461)

---
 python/sglang/srt/layers/moe/topk.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/python/sglang/srt/layers/moe/topk.py b/python/sglang/srt/layers/moe/topk.py
index f17b908cb..e59f0e299 100644
--- a/python/sglang/srt/layers/moe/topk.py
+++ b/python/sglang/srt/layers/moe/topk.py
@@ -220,6 +220,8 @@ def biased_grouped_topk(
     # TODO: moe_fused_gate kernel is not supported for n_share_experts_fusion > 0 now.
     if (
         _is_cuda
+        and gating_output.shape[1] // num_expert_group
+        <= 32  # moe_fused_gate kernel ensure 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 n_share_experts_fusion == 0
         and is_power_of_two(correction_bias.shape[0])
     ):