Very bad interview. The interviewer expected you to have vLLM experience, otherwise they interrupted your self-introduction and went through the questions. I could not answer fully correctly as I do not have any vLLM experience before, then the interviewer said it was a waste of both time.

1. Imagine one customer has one small model host in one GPU, they deploy in other company then it can hit 100 RPS while it deploy in Nebius only can hit 50 RPS, how to do optimisation. You need to ask then he will tell you that the avg input token 300 output token 300, no KV cache iteration.

2. Customer has large LLM model deploy on 4 replicas, each host 1 H100 GPU. Input token 50k token, output ~100 token. What's the optimisation?