Part II — Inside a Single Engine

One request, one engine, opened end to end. A scheduler chooses the next unit of work, a memory manager finds room for its state, kernels move the bytes, compiled graphs remove the launch cost, and the numerical and decoding choices determine what the engine may safely trade away. Chapter 12 extends the single-engine story to multi-tenant adapter serving, where many customized models share one set of base weights.

Chapters 5–12