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

Fuse routed_scaling_factor in DeepSeek (#6710)

Browse Source
This commit is contained in:
fzyzcjy
2025-05-29 06:53:37 +08:00
committed by GitHub
parent 5170b010a6
commit 541a985f85
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/srt/models/deepseek_v2.py
View File

@@ -526,9 +526,13 @@ class DeepseekV2MoE(nn.Module):
def op_output(self, state):
final_hidden_states = state.pop("hidden_states_after_combine")
final_hidden_states *= self.routed_scaling_factor
if (s := state.pop("shared_output")) is not None:
final_hidden_states = final_hidden_states + s
if (shared_output := state.pop("shared_output")) is not None:
x = shared_output
x.add_(final_hidden_states, alpha=self.routed_scaling_factor)
final_hidden_states = x
else:
final_hidden_states *= self.routed_scaling_factor
state.hidden_states_mlp_output = final_hidden_states