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

[LongCat] Optimize zero_experts_compute_triton by changing mask (#10303)

Browse Source
This commit is contained in:
zk-lover
2025-09-12 05:56:25 +08:00
committed by GitHub
parent 46ccbed2cd
commit c5d2b01cea
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/layers/moe/ep_moe/kernels.py
View File

@@ -1416,7 +1416,7 @@ def zero_experts_compute_triton(
zero_expert_scales[zero_expert_mask] = 0.0
normal_expert_mask = expert_indices >= num_experts
expert_indices[normal_expert_mask] = 0
expert_indices[normal_expert_mask] = -1
expert_scales[normal_expert_mask] = 0.0
output = torch.zeros_like(hidden_states).to(hidden_states.device)