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

Speedup shared expert weight construction by avoid cloning (#5188)

Browse Source
This commit is contained in:
fzyzcjy
2025-04-21 09:12:01 +08:00
committed by GitHub
parent f08154193c
commit 5239d79568
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/models/deepseek_v2.py
View File

@@ -1628,7 +1628,7 @@ class DeepseekV2ForCausalLM(nn.Module):
f"mlp.experts." f"mlp.experts."
f"{self.config.n_routed_experts + num_repeat}" f"{self.config.n_routed_experts + num_repeat}"
f".{suffix}", f".{suffix}",
weights_dict[shared_expert_weight_name].clone(), weights_dict[shared_expert_weight_name],
) )
) )
names_to_remove += [shared_expert_weight_name] names_to_remove += [shared_expert_weight_name]