[AutoDeploy]: Qwen-3.5 400B perf optimization

[galagam]

🚀 The feature, motivation and pitch

Follow up on #11440
BS=512, 1000/1000, B200, TP=8

Output TPS ~0.5x, TTFT is ~0.2x compared to vLLM.
Targeting >0.8x as initial target.

Dump traces and identify performance optimizations.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

[taylor-yb-lee]

https://docs.google.com/document/d/1swpR_MhUknXsQ6d5waVQengFKpdGBeSpvkI1kKKYv94/edit?tab=t.0 Updating daily status & progress here