Distributed and Parallel Training Tutorials

Distributed training is a model training paradigm that involves spreading training workload across multiple worker nodes, therefore significantly improving the speed of training and model accuracy. While distributed training can be used for any type of ML model training, it is most beneficial to use it for large models and compute demanding tasks as deep learning.

There are a few ways you can perform distributed training in PyTorch with each method having their advantages in certain use cases:

• DistributedDataParallel (DDP)

• Fully Sharded Data Parallel (FSDP)

• Tensor Parallel (TP)

• Device Mesh

• Remote Procedure Call (RPC) distributed training

• Custom Extensions

Read more about these options in Distributed Overview.

Learn DDP

DDP Intro Video Tutorials

A step-by-step video series on how to get started with DistributedDataParallel and advance to more complex topics

Code Video

Getting Started with Distributed Data Parallel

This tutorial provides a short and gentle intro to the PyTorch DistributedData Parallel.

Code

Distributed Training with Uneven Inputs Using the Join Context Manager

This tutorial describes the Join context manager and demonstrates it’s use with DistributedData Parallel.

Code

Learn FSDP

Getting Started with FSDP

This tutorial demonstrates how you can perform distributed training with FSDP on a MNIST dataset.

Code

FSDP Advanced

In this tutorial, you will learn how to fine-tune a HuggingFace (HF) T5 model with FSDP for text summarization.

Code

Learn Tensor Parallel (TP)

Large Scale Transformer model training with Tensor Parallel (TP)

This tutorial demonstrates how to train a large Transformer-like model across hundreds to thousands of GPUs using Tensor Parallel and Fully Sharded Data Parallel.

Code

Learn DeviceMesh

Getting Started with DeviceMesh

In this tutorial you will learn about DeviceMesh and how it can help with distributed training.

Code

Learn RPC

Getting Started with Distributed RPC Framework

This tutorial demonstrates how to get started with RPC-based distributed training.

Code

Implementing a Parameter Server Using Distributed RPC Framework

This tutorial walks you through a simple example of implementing a parameter server using PyTorch’s Distributed RPC framework.

Code

Implementing Batch RPC Processing Using Asynchronous Executions

In this tutorial you will build batch-processing RPC applications with the @rpc.functions.async_execution decorator.

Code

Combining Distributed DataParallel with Distributed RPC Framework

In this tutorial you will learn how to combine distributed data parallelism with distributed model parallelism.

Code

Custom Extensions

Customize Process Group Backends Using Cpp Extensions

In this tutorial you will learn to implement a custom ProcessGroup backend and plug that into PyTorch distributed package using cpp extensions.

Code