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

[DeepEP] Reduce routed scaling overhead (#5277)

Browse Source 
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
This commit is contained in:
yulei
2025-04-14 07:03:09 +08:00
committed by GitHub
parent 39d90449f3
commit adca585bfb
1 changed files with 9 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/models/deepseek_v2.py
View File

@@ -337,8 +337,7 @@ class DeepseekV2MoE(nn.Module):
topk_weights, topk_weights,
forward_mode=forward_mode, forward_mode=forward_mode,
) )
final_hidden_states = ( final_hidden_states = self.experts(
self.experts(
hidden_states=hidden_states, hidden_states=hidden_states,
reorder_topk_ids=reorder_topk_ids, reorder_topk_ids=reorder_topk_ids,
seg_indptr=seg_indptr, seg_indptr=seg_indptr,
@@ -346,8 +345,6 @@ class DeepseekV2MoE(nn.Module):
expected_m=expected_m, expected_m=expected_m,
forward_mode=forward_mode, forward_mode=forward_mode,
) )
* self.routed_scaling_factor
)
if self.ep_size > 1: if self.ep_size > 1:
final_hidden_states = self.deepep_dispatcher.combine( final_hidden_states = self.deepep_dispatcher.combine(
final_hidden_states, final_hidden_states,
@@ -355,6 +352,8 @@ class DeepseekV2MoE(nn.Module):
topk_weights, topk_weights,
forward_mode, forward_mode,
) )
final_hidden_states *= self.routed_scaling_factor
if shared_output is not None: if shared_output is not None:
final_hidden_states = final_hidden_states + shared_output final_hidden_states = final_hidden_states + shared_output