CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 11: Inference and Serving · part 82 of 111
Part 82 · Module 11: Inference and Serving

Lesson 5: Quantisation: fewer bits per weight, measured

29 min read·13 Oct 2026
Module 11 · Lesson 5 · About 60 minutes · Needs: Module 10, Lesson 3 (fake_quant and KV cache quantisation); Module 6's base recipe; Module 8's package as pasted in Module 10, Lesson 3

The situation

Sol 434 at Aurora Station, the second day of the Earth link outage. Hana Aoki, the Compute Officer, wants a copy of base on each robot's controller, and those have very little memory. "Can it live in a sixth of its size? And what does it lose?"

You follow one log header, Log 4821, sol 212, the night shift., and one row of weights, at 8, 4, 3 and 2 bits. Module 10, Lesson 3 shrank the KV cache; this lesson shrinks weights and activations.

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.