Tinker Mixture of experts

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

Loading Larkspur’s toy experts model…

“Every enterprise account has a named customer success”

Chapter 6 of 7

Total versus active

Back to the vendor’s card: “47B total, 13B active”. Larkspur needs to know what each of the two numbers costs.

This chapter: What does the total cost, and what does the active count cost?

About this chapter

You will learn

  • Total weights set the memory
  • Active weights set the work per token
  • What the toy measured against a dense model

Words you’ll meet

Total parameters
Every weight in the model, every expert included. All of them must sit in memory.
Active parameters
The weights one token runs through: shared parts plus its chosen experts.
Memory
The GPU memory that holds the weights while the model serves.

Step 1 of 5

Every room’s weights wait in , ready, but each token works through only two rooms a floor.

Count the toy’s weights.