A high-throughput and memory-efficient inference and serving engine for LLMs
[Bugfix] Fix nemotron accuracy drop introduced by #41184 (#45037)
Signed-off-by: Bill Nell <bnell@redhat.com>
bnellnm committed
Jun 10, 2026 at 13:16 UTC
29026682cb29335dcea15d82c14ee18fff40f71d
1 file changed
+3
-1
vllm/model_executor/layers/fused_moe/layer.py
+3
-1
@@ -349,7 +349,9 @@ def FusedMoE(
349
topk_group=topk_group,
350
custom_routing_function=custom_routing_function,
351
scoring_func=scoring_func,
352
- routed_scaling_factor=routed_scaling_factor,
352
+ routed_scaling_factor=routed_scaling_factor
353
+ if not apply_routed_scale_to_output
354
+ else 1.0,
355
swiglu_limit=swiglu_limit,
356
# TODO get from router? needs to be truncated?
357
e_score_correction_bias=e_score_correction_bias,