…er if epmoe is enabled

Motivation

Issue: #8402

Also noticed significant regression when running epmoe during recent GLM4.5 support work: GSM8K accuracy drops from 0.965 to 0.745 when EPMOE is enabled. Accuracy is good for TP & DeepEP.

python3 -m sglang.launch_server --model  /shared/public/elr-models/zai-org/GLM-4.5 --tp-size 8 --trust-remote-code

python3 benchmark/gsm8k/bench_sglang.py

Modifications

There is a bug in DeepSeek-V2 and GLM-4.5 related to how routed_scaling_factor is applied in MoE (Mixture-of-Experts) layers.

Currently, the routed_scaling_factor is applied in three different places, leading to ambiguity:

• self.topk.forward: The scaling is applied only in the n out of m condition, but not in the m - n out of m condition.
• self.experts.forward: Same as above — the factor is only applied in n out of m.
• Model-level logic: The model itself may apply routed_scaling_factor, but it cannot know whether self.topk or self.experts have already done so.

This results in uncertain and inconsistent scaling, as the model layer has no visibility into whether routed_scaling_factor has already been applied upstream.

🎪 TL;DR: The model can't know if it should apply * routed_scaling_factor or not, because topk and experts may or may not have already done it, depending on the code path.

My PR forces model layer to apply * routed_scaling_factor when EPMOE is enabled because from the current codebase, epmoe won't apply *routed_scaling_factor by itself.

Need to follow up and possibly refactor sglang moe codebase to make it clear which layer should apply * routed_scaling_factor

Accuracy Test

• GLM4.5 GSM8K accuracy jumps from 0.745 to 0.965 under EPMOE which is the same as TP & DeepEP results
• Need to run gsm8k for deepseek-v3 too - WIP

Benchmark & Profiling

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.