Serving AI models at scale with vLLM