Given that i have a rate estimate of K requests per minute, a given hardware and a model running on it using VLLM, how can i get the best performance?