Tinker Diffusion language models

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

The phrase: “ safety boots and high-visibility”

Chapter 1 of 7

One word at a time

A vendor promises a model that writes 1,009 tokens a second. Larkspur’s team wants to know what that speed buys.

This chapter: How does a model write a reply, and what does each token cost?

About this chapter

You will learn

  • Why a left-to-right model takes one pass per token
  • What a pass is
  • Why passes set a model’s speed

Words you’ll meet

Token
A piece of text a model reads and writes: a word, part of a word, or a mark.
Pass
One run of the model over everything it can see. Each pass takes about the same time.
Larkspur mini
Larkspur’s tiny left-to-right model from the Transformers module. It writes one token a pass.

Step 1 of 3

, the typewriter, continues a phrase from the safety handbook. Each is one run of the model.

Predict first.

Larkspur mini must write 16 tokens. How many passes does it take?