Part III: Training and single-worker inference

When included, the optional training lab collection is in the Optional Labs appendix. The core training update is complete on its own; the lab adds a longer profiling experiment.

Four chapter scenes compare a training forward, loss, backward and optimizer update; prompt prefill followed by one decode call with separate token and cache paths; INT8 code storage and fused conversion/compute with separate result checks; and current-token Q/K/V projections beside a logical-to-physical KV block table. Memory panels distinguish fixed weights, growing request KV, temporary workspace and runtime headroom.
Figure 1: Training updates persistent state; generation keeps weights fixed while KV grows. Precision, compression and cache layout change storage and computation in different ways.

Chapters in this part