vllm

[Perf] Optimize async scheduling redundant copy, 0.9% E2E throughput improvement

#34029

Open