CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 11: Inference and Serving · part 84 of 111
Part 84 · Module 11: Inference and Serving

Lesson 7: Serving in practice: quality, engines, adapters, cost, honest numbers

32 min read·13 Oct 2026
Module 11 · Lesson 7 · About 50 minutes · Needs: Lesson 1 of this module (the serving contract and the cached decoder); Lessons 3 to 6 help but are not required

The situation

Sol 436, the last day of the outage. Hana Aoki, the Compute Officer, wants one page before the link returns: which compressed base to keep, what an engine adds, whether one server can hold Maya's planned variants, and what a token costs. "Every number on it must survive Kofi."

The running example is Tariq's request from Lesson 1, Log 4821, sol 434, the night shift., continued greedily: "Maya Nilsson, robotics engineer, replaced the airlock motor".

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.