vllm

[SpecDecode] Reduce TP communication for large-vocab draft models speculative decoding

#39419

Merged