UCSD CSE 291A Lecture 13 covers prefix caching and P–D disaggregation

UCSD CSE 291A Lecture 13, taught by Yiying Zhang, covers prefix caching and Prefill–Decode (P–D) disaggregation for distributed LLM serving, according to a post by meng shao on X. The main reading is Preble (ICLR 2025), whose global E2 scheduler and local fair scheduler report 1.5–14.5x lower average latency than SGLang in the lecture's own results. The lecture also covers DistServe, which splits prefill and decode workers and reports about 3.3 requests/s per GPU in an illustrative example.