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

Remve router gemm output dtype conversion (#8204)

Browse Source
This commit is contained in:
Ke Bao
2025-07-21 15:37:00 +08:00
committed by GitHub
parent 9b5de6cb06
commit 6936be3221
1 changed files with 2 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/models/deepseek_v2.py
View File

@@ -254,9 +254,8 @@ class MoEGate(nn.Module):
and self.weight.shape[0] == 256 and self.weight.shape[0] == 256
and _device_sm >= 90 and _device_sm >= 90
): ):
logits = dsv3_router_gemm(hidden_states, self.weight).to( # router gemm output float32
hidden_states.dtype logits = dsv3_router_gemm(hidden_states, self.weight)
)
else: else:
logits = F.linear(hidden_states, self.weight, None) logits = F.linear(hidden_states, self.weight, None)