vllm

fix(spec_decode): ensure draft model uses correct parallel/model config in all proposers

#44086

Open