vllm

[Perf] Optimize `sampled_token_ids` using numpy and remove `tolist`, 0.9% E2E throughput improvement

#35446

Open