The Finished Engine

ch29 · The Finished Engine [DRAFT]¶

The problem¶

No problem left to solve — this chapter accounts for what was solved.

[To write: open with the measurement from the previous chapter that is unacceptable. Show it, do not assert it.]

The idea¶

• The full scorecard: baseline to finished engine, every row, with the chapter that moved it

• Design retrospective: which changes paid for their complexity and which did not

• What we deliberately did not build — multi-node, MoE routing at scale, custom CUDA beyond ch13, disaggregation at real scale — and what each omission costs

• A guided map into the real codebases: where each concept in this book lives in vLLM, SGLang and TensorRT-LLM, so the reader can now read them fluently

[To write: develop the mechanism from first principles, with the arithmetic that predicts the result.]

The build¶

• Nothing new. Consolidate, tag the final checkpoint, and point outward

[To write: incremental implementation. Quote code from llmserve/ with {literalinclude} and :start-at: / :end-at: anchors — never paste it inline.]

The measurement¶

The complete table, plus the ratio between the first and last rows.

[To write: re-run the harness, append the scorecard row, and explain in one paragraph why the number moved. Read the figures from bench/results/ — never type a number into prose (PLAN.md §6.3).]

The cost¶

The engine is a teaching artefact, not production software. State clearly what a reader should use in production instead, and why that is the right call.

[To write: what did this make worse? Complexity, latency variance, quality, memory, operational burden. This section is mandatory — the chapter is not done without it (PLAN.md §12.3).]

Key takeaways¶

• [To write: 3–5 bullets, each a claim the chapter earned.]

Looking ahead¶

[To write: the transition that sets up the next chapter’s problem.]

Further reading¶

[To write: primary sources, cited with MyST citation syntax against references.bib.]