We study how to approximate the famous shortest-job-first scheduling in LLM inference!