vllm

[Optimization] Cache sampled token ids in model runner

#20291

Merged