Part IV: Serving systems

When included, the optional decode and serving labs are in the Optional Labs appendix. They collect longer cache, trace and decoding probes without supplying setup required by the main chapters.

Three chapter scenes show request admission, changing batches and shared prefix KV; draft verification with rejection correction and grammar-constrained token choice; and client timing, offered-load latency, fair comparisons and serving-stack roles.
Figure 1: Request scheduling manages shared capacity, decoding methods change token production under their own rules, and client measurements show the resulting service behavior.

Chapters in this part