A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[BugFix] Fix routed_scaling_factor double mul for dots1 and glm4 MoE models (#24132)
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
Yong Hoon Shin committed
Sep 2, 2025 at 21:57 UTC
426cc8629f7e630e1c5a0b96fe2db737a170a06d
2 files changed
+4
-2
vllm/model_executor/models/dots1.py
+2
-1
@@ -137,7 +137,8 @@ class Dots1MoE(nn.Module):
137
topk_group=config.topk_group,
138
prefix=f"{prefix}.experts",
139
scoring_func=config.scoring_func,
140
- routed_scaling_factor=self.routed_scaling_factor,
140
+ # we do scaling outside, set factor to 1.0 to avoid double mul
141
+ routed_scaling_factor=1.0,
142
e_score_correction_bias=self.gate.e_score_correction_bias)
143
144
if config.n_shared_experts is not None:
vllm/model_executor/models/glm4_moe.py
+2
-1
@@ -159,7 +159,8 @@ class Glm4MoE(nn.Module):
159
topk_group=config.topk_group,
160
prefix=f"{prefix}.experts",
161
scoring_func="sigmoid",
162
- routed_scaling_factor=self.routed_scaling_factor,
162
+ # we do scaling outside, set factor to 1.0 to avoid double mul
163
+ routed_scaling_factor=1.0,
164
e_score_correction_bias=self.gate.e_score_correction_bias,
165
enable_eplb=self.enable_eplb,
166
num_redundant_experts=self.n_redundant_experts)