LLMs served with vLLM behind llm-d's router, which sends each request to the replica that can answer it fastest.