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

Fix biased_grouped_topk_cpu (#9420)

Browse Source
This commit is contained in:
Cao E
2025-08-21 10:18:48 +08:00
committed by GitHub
parent af1973b871
commit c674bf9c6b
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

@@ -709,8 +709,10 @@ def biased_grouped_topk_cpu(
routed_scaling_factor: Optional[float] = None, routed_scaling_factor: Optional[float] = None,
num_token_non_padded: Optional[torch.Tensor] = None, num_token_non_padded: Optional[torch.Tensor] = None,
expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None, expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
apply_routed_scaling_factor_on_output: Optional[bool] = False,
): ):
assert expert_location_dispatch_info is None assert expert_location_dispatch_info is None
assert not apply_routed_scaling_factor_on_output, "Not implemented"
return torch.ops.sgl_kernel.biased_grouped_topk_cpu( return torch.ops.sgl_kernel.biased_grouped_topk_cpu(
hidden_states, hidden_states,
gating_output, gating_output,