vllm

[Bugfix][MoE] Fix 6-8% decode regression: prefer multi-stream shared expert overlap

#38990

Merged