How we get sub-second cold starts and serve multiple models on one GPU with vLLM