A Hybrid Tensor-Expert-Data Parallelism Approach to Optimize Mixture-of-Experts Training

TL;DR

This paper introduces DeepSpeed-TED, a hybrid parallelism method combining data, tensor, and expert parallelism to efficiently train large-scale MoE models, achieving significant speedups and enabling larger models than previous methods.

Contribution

The paper presents a novel three-dimensional hybrid parallel algorithm for MoE training, including memory and communication optimizations, allowing training of models 4-8 times larger than current state-of-the-art.

Findings

Achieved 26% speedup over baseline training.

Enabled training of 40 billion parameter MoE models.

Implemented in DeepSpeed with effective communication optimizations.

Abstract

Mixture-of-Experts (MoE) is a neural network architecture that adds sparsely activated expert blocks to a base model, increasing the number of parameters without impacting computational costs. However, current distributed deep learning frameworks are limited in their ability to train high-quality MoE models with large base models. In this work, we present DeepSpeed-TED, a novel, three-dimensional, hybrid parallel algorithm that combines data, tensor, and expert parallelism to enable the training of MoE models with 4 to 8x larger base models than the current state-of-the-art. We also describe memory optimizations in the optimizer step, and communication optimizations that eliminate unnecessary data movement. We implement our approach in DeepSpeed and achieve speedups of 26% over a baseline (i.e. without our communication optimizations) when training a 40 billion parameter MoE model (6.7 billion base model with 16 experts) on 128 V100 GPUs.