Hyperparameter Transfer with Mixture-of-Expert Layers

TL;DR

This paper introduces a new parameterization for transformer models with Mixture-of-Experts layers, enabling reliable hyperparameter transfer across different model sizes and training conditions.

Contribution

It proposes a novel parameterization justified by dynamical mean-field theory, facilitating hyperparameter transfer in MoE models across various scales.

Findings

Hyperparameter transfer is reliable across models from 51M to 2B parameters.

The new parameterization enables effective scaling of MoE models.

Hyperparameters from small models can be used to train larger models efficiently.

Abstract

Mixture-of-Experts (MoE) layers have emerged as an important tool in scaling up modern neural networks by decoupling total trainable parameters from activated parameters in the forward pass for each token. However, sparse MoEs add complexity to training due to (i) new trainable parameters (router weights) that, like all other parameter groups, require hyperparameter (HP) tuning; (ii) new architecture scale dimensions (number of and size of experts) that must be chosen and potentially taken large. To make HP selection cheap and reliable, we propose a new parameterization for transformer models with MoE layers when scaling model width, depth, number of experts, and expert (hidden) size. Our parameterization is justified by a novel dynamical mean-field theory (DMFT) analysis. When varying different model dimensions trained at a fixed token budget, we find empirically that our parameterization enables reliable HP transfer across models from 51M to over 2B total parameters. We further take HPs identified from sweeping small models on a short token horizon to train larger models on longer horizons and report performant model behaviors.