xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-07 15:35:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] Fix routed_scaling_factor double mul for dots1 and glm4 MoE models (#24132)

Browse Source 
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
This commit is contained in:
Yong Hoon Shin 2025-09-02 21:57:59 -07:00 committed by GitHub
parent e81d4e69c1
commit 426cc8629f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/model_executor/models/dots1.py
View File

@ -137,7 +137,8 @@ class Dots1MoE(nn.Module):
topk_group=config.topk_group,
prefix=f"{prefix}.experts",
scoring_func=config.scoring_func,
routed_scaling_factor=self.routed_scaling_factor,
# we do scaling outside, set factor to 1.0 to avoid double mul
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias)
if config.n_shared_experts is not None:

3
vllm/model_executor/models/glm4_moe.py
View File

@ -159,7 +159,8 @@ class Glm4MoE(nn.Module):
topk_group=config.topk_group,
prefix=f"{prefix}.experts",
scoring_func="sigmoid",
routed_scaling_factor=self.routed_scaling_factor,
# we do scaling outside, set factor to 1.0 to avoid double mul
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias,
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts)