Serving a 1B math model on vLLM: the memory cap sets how many answers fit, not how fast