Chapter 17
17Serving Systems in Production
17.1What a serving stack must do
Planned: Throughput, latency (TTFT and TPOT), and cost as competing goals.
17.2The engines
Planned: vLLM, TensorRT-LLM, SGLang, TGI; what differentiates them.
17.3Multi-request scheduling
Planned: Prefix caching, prioritization, and fairness across users.
17.4Autoscaling and the economics
Planned: GPU utilization, cold starts, and the unit economics of tokens.