Description

This PR introduces a small, inference specific optimization to reduce overall step time. Specifically, this PR moves the per-expert scale application to the model initialization instead of during the call method, given that these parameters are static during inference.

Tests

vllm_decode.py before and after.

Step time without this PR: 13.59 ms
Step time with this PR: 7.975 ms

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.