EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

fix kimi vl running bug after rebase main (#5461)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-04-18 15:17:34 +08:00
committed by GitHub
parent b2a189dd11
commit bed05878f6
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/layers/moe/topk.py
View File

@@ -220,6 +220,8 @@ def biased_grouped_topk(
# TODO: moe_fused_gate kernel is not supported for n_share_experts_fusion > 0 now. # TODO: moe_fused_gate kernel is not supported for n_share_experts_fusion > 0 now.
if ( if (
_is_cuda _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 n_share_experts_fusion == 0
and is_power_of_two(correction_bias.shape[0]) and is_power_of_two(correction_bias.shape[0])
): ):