vllm

[Perf] Optimize compute maxsim using batched version, 3.2% E2E throughput improvement

#36710

Merged