Tinker Mixture of experts

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

Loading Larkspur’s toy experts model…

Pick a token

“We reply within five business”

Chapter 2 of 7

Rooms and a doorman

The vendor’s model splits its big kitchen into rooms. A doorman at each floor sends every token to a few of them.

This chapter: Who picks the rooms for a token, and how many?

About this chapter

You will learn

  • What an expert is
  • How the router scores every expert
  • Why only the top two run

Words you’ll meet

Expert
One of the rooms: a small MLP of its own. Each layer here has eight.
Router
The doorman: a small layer that scores every expert for each token.
Top-2
Only the two best-scored experts run; their outputs mix by the router’s scores.

Step 1 of 5

Each floor holds 8 rooms, the . Its doorman, the , scores every room for each token.

Predict first.

The doorman scores all 8 rooms for a token. Where does the token go?