minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[NPU] Acc fix for afmoe model introduced by topk refactor. (#31280)

Browse Source
This commit is contained in:
McZyWu
2026-07-27 15:40:49 +08:00
committed by GitHub
parent c0f47a06fc
commit 169fc1e20c
1 changed files with 9 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/afmoe.py
+9 -4
View File
@@ -233,9 +233,7 @@ class AfmoeMoE(nn.Module):
expert_bias=self.expert_bias, expert_bias=self.expert_bias,
) )
renormalize = ( renormalize = self.route_norm if self.score_func == "sigmoid" else False
self.route_norm if self.score_func == "sigmoid" and not _is_npu else False
)
self.topk = TopK( self.topk = TopK(
top_k=self.top_k, top_k=self.top_k,
renormalize=renormalize, renormalize=renormalize,
@@ -245,7 +243,14 @@ class AfmoeMoE(nn.Module):
custom_routing_function=custom_routing_fn, custom_routing_function=custom_routing_fn,
correction_bias=correction_bias, correction_bias=correction_bias,
routed_scaling_factor=self.route_scale, routed_scaling_factor=self.route_scale,
**({"scoring_func": self.score_func} if _is_npu else {}), **(
{
"scoring_func": self.score_func,
"apply_routed_scaling_factor_on_output": True,
}
if _is_npu
else {}
),
) )
def pack_params(self) -> None: def pack_params(self) -> None: