FastMoE: A Fast Mixture-of-Expert Training System

TL;DR

FastMoE is a high-performance, open-source distributed MoE training system built on PyTorch, enabling scalable, efficient training of trillion-parameter language models across multiple GPUs and nodes.

Contribution

It introduces a flexible, GPU-compatible MoE training system with optimized acceleration techniques and hierarchical interfaces for model design and adaptation.

Findings

Supports linear scaling of experts with GPUs

Enables training of trillion-parameter models

Optimized for high-performance distributed training

Abstract

Mixture-of-Expert (MoE) presents a strong potential in enlarging the size of language model to trillions of parameters. However, training trillion-scale MoE requires algorithm and system co-design for a well-tuned high performance distributed training system. Unfortunately, the only existing platform that meets the requirements strongly depends on Google's hardware (TPU) and software (Mesh Tensorflow) stack, and is not open and available to the public, especially GPU and PyTorch communities. In this paper, we present FastMoE, a distributed MoE training system based on PyTorch with common accelerators. The system provides a hierarchical interface for both flexible model design and easy adaption to different applications, such as Transformer-XL and Megatron-LM. Different from direct implementation of MoE models using PyTorch, the training speed is highly optimized in FastMoE by sophisticated high-performance acceleration skills. The system supports placing different experts on multiple GPUs across multiple nodes, enabling enlarging the number of experts linearly against the number of GPUs. The source of FastMoE is available at https://github.com/laekov/fastmoe under Apache-2 license.