vllm

[torch.compile] Add an option to force-enable the MOE cold start optimization

#33735

Merged