Three types of LLM workloads and how to serve them