Awesome

Introduction

An easy-to-use and efficient system to support the Mixture of Experts (MoE) model for PyTorch.

Installation

Prerequisites

PyTorch with CUDA is required. The repository is currently tested with PyTorch v1.10.0 and CUDA 11.3, with designed compatibility to older and newer versions.

The minimum version of supported PyTorch is 1.7.2 with CUDA 10. However, there are a few known issues that requires manual modification of FastMoE's code with specific older dependents.

If the distributed expert feature is enabled, NCCL with P2P communication support, typically versions >=2.7.5, is needed.

Installing

FastMoE contains a set of PyTorch customized opearators, including both C and Python components. Use python setup.py install to easily install and enjoy using FastMoE for training.

A step-by-step tutorial for the installation procedure can be found here.

The distributed expert feature is enabled by default. If you want to disable it, pass environment variable USE_NCCL=0 to the setup script.

Note that an extra NCCL developer package is needed, which has to be consistent with your PyTorch's NCCL version, which can be inspected by running torch.cuda.nccl.version(). The official PyTorch docker image is recommended, as the environment is well-setup there. Otherwise, you can access the download link of all NCCL versions to download the NCCL package that is suitable for you.

Usage

FMoEfy a Transformer model

Transformer is currently one of the most popular models to be extended by MoE. Using FastMoE, a Transformer-based model can be extended as MoE by an one-key plugin shown as follow.

For example, when using Megatron-LM, using the following lines can help you easily scale up the MLP layers to multiple experts.

model = ...

from fmoe.megatron import fmoefy
model = fmoefy(model, fmoe_num_experts=<number of experts per worker>)

train(model, ...)

A detailed tutorial to moefy Megatron-LM can be found here.

Using FastMoE as a PyTorch module

An example MoE transformer model can be seen in the Transformer-XL example. The easist way is to replace the MLP layer by the FMoE layers.

Using FastMoE in Parallel

FastMoE supports multiple ways of parallel training. See a comprehensive document for parallelism for details. Below shows the two simplest ways of using FastMoE in parallel.

Data Parallel

In FastMoE's data parallel mode, both the gate and the experts are replicated on each worker. The following figure shows the forward pass of a 3-expert MoE with 2-way data parallel.

For data parallel, no extra coding is needed. FastMoE works seamlessly with PyTorch's DataParallel or DistributedDataParallel. The only drawback of data parallel is that the number of experts is constrained by each worker's memory.

Expert Parallel (also called Model Parlallel in some previous versions)

In FastMoE's expert parallel mode, the gate network is still replicated on each worker but experts are placed separately across workers. Thus, by introducing additional communication cost, FastMoE enjoys a large expert pool whose size is proportional to the number of workers.

The following figure shows the forward pass of a 6-expert MoE with 2-way model parallel. Note that experts 1-3 are located in worker 1 while experts 4-6 are located in worker 2.

FastMoE's expert parallel requires sophiscated parallel strategies that neither PyTorch nor Megatron-LM provided when FastMoE was created. The fmoe.DistributedGroupedDataParallel module is introduced to replace PyTorch's DDP module.

Faster Performance Features

From a PPoPP'22 paper, FasterMoE: modeling and optimizing training of large-scale dynamic pre-trained models, we have adopted techniques to make FastMoE's model parallel much more efficient.

These optimizations are named as Faster Performance Features, and can be enabled via several environment variables. Their usage and constraints are detailed in a separate document.

Citation

For the core FastMoE system.

@article{he2021fastmoe,
      title={FastMoE: A Fast Mixture-of-Expert Training System}, 
      author={Jiaao He and Jiezhong Qiu and Aohan Zeng and Zhilin Yang and Jidong Zhai and Jie Tang},
      journal={arXiv preprint arXiv:2103.13262},
      year={2021}
}

For the faster performance features.

@inproceedings{he2022fastermoe,
    author = {He, Jiaao and Zhai, Jidong and Antunes, Tiago and Wang, Haojie and Luo, Fuwen and Shi, Shangfeng and Li, Qin},
    title = {FasterMoE: Modeling and Optimizing Training of Large-Scale Dynamic Pre-Trained Models},
    year = {2022},
    isbn = {9781450392044},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3503221.3508418},
    doi = {10.1145/3503221.3508418},
    booktitle = {Proceedings of the 27th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming},
    pages = {120–134},
    numpages = {15},
    keywords = {parallelism, distributed deep learning, performance modeling},
    location = {Seoul, Republic of Korea},
    series = {PPoPP '22}
}

Troubleshootings / Discussion

If you have any problem using FastMoE, or you are interested in getting involved in developing FastMoE, feel free to join our slack channel.