Distributed Training Made Easy with PyTorch-Ignite

Writing agnostic distributed code that supports different platforms, hardware configurations (GPUs, TPUs) and communication frameworks is tedious. In this blog, we will discuss how PyTorch-Ignite solves this problem with minimal code change.

Crossposted from https://labs.quansight.org/blog/2021/06/distributed-made-easy-with-ignite/

Prerequisites

This blog assumes you have some knowledge about:

1. PyTorch’s distributed package, the backends and collective functions it provides. In this blog, we will focus on distributed data parallel code.

2. PyTorch-Ignite. Refer to this blog for a quick high-level overview.

Introduction

PyTorch-Ignite’s ignite.distributed (idist) submodule introduced in version v0.4.0 (July 2020) quickly turns single-process code into its data distributed version.

Thus, you will now be able to run the same version of the code across all supported backends seamlessly:

• backends from native torch distributed configuration: nccl, gloo, mpi.

• Horovod framework with gloo or nccl communication backend.

• XLA on TPUs via pytorch/xla.

In this blog post we will compare PyTorch-Ignite’s API with torch native’s distributed code and highlight the differences and ease of use of the former. We will also show how Ignite’s auto_* methods automatically make your code compatible with the aforementioned distributed backends so that you only have to bring your own model, optimizer and data loader objects.

Code snippets, as well as commands for running all the scripts, are provided in a separate repository.

Then we will also cover several ways of spawning processes via torch native torch.multiprocessing.spawn and also via multiple distributed launchers in order to highlight how Pytorch-Ignite’s idist can handle it without any changes to the code, in particular:

• torch.multiprocessing.spawn

• torchrun

• horovodrun

• slurm

More information on launchers experiments can be found here.

PyTorch-Ignite Unified Distributed API

We need to write different code for different distributed backends. This can be tedious especially if you would like to run your code on different hardware configurations. Pytorch-Ignite’s idist will do all the work for you, owing to the high-level helper methods.

Focus on the helper auto_* methods

• auto_model()

This method adapts the logic for non-distributed and available distributed configurations. Here are the equivalent code snippets for distributed model instantiation:

Additionally, it is also compatible with NVIDIA/apex

model, optimizer = amp.initialize(model, optimizer, opt_level=opt_level)
model = idist.auto_model(model)

and Torch native AMP

model = idist.auto_model(model)

with autocast():
    y_pred = model(x)

• auto_optim()

This method adapts the optimizer logic for non-distributed and available distributed configurations seamlessly. Here are the equivalent code snippets for distributed optimizer instantiation:

• auto_dataloader()

This method adapts the data loading logic for non-distributed and available distributed configurations seamlessly on target devices.

Additionally, auto_dataloader() automatically scales the batch size according to the distributed configuration context resulting in a general way of loading sample batches on multiple devices.

Here are the equivalent code snippets for the distributed data loading step:

Examples

The code snippets below highlight the API’s specificities of each of the distributed backends on the same use case as compared to the idist API. PyTorch native code is available for DDP, Horovod, and for XLA/TPU devices.

PyTorch-Ignite’s unified code snippet can be run with the standard PyTorch backends like gloo and nccl and also with Horovod and XLA for TPU devices. Note that the code is less verbose, however, the user still has full control of the training loop.

The following examples are introductory. For a more robust, production-grade example that uses PyTorch-Ignite, refer here.

The complete source code of these experiments can be found here.

PyTorch-Ignite - Torch native Distributed Data Parallel - Horovod - XLA/TPUs

Horovod	Torch XLA
Source Code	Source Code

dist.init_process_group(backend, store=..., world_size=world_size, rank=rank)
rank = idist.get_rank()
ws = idist.get_world_size()
model = idist.auto_model(model)
dist.destroy_process_group()

Running Distributed Code

PyTorch-Ignite’s idist also unifies the distributed codes launching method and makes the distributed configuration setup easier with the ignite.distributed.launcher.Parallel (idist Parallel) context manager.

This context manager has the capability to either spawn nproc_per_node (passed as a script argument) child processes and initialize a processing group according to the provided backend or use tools like torchrun, slurm, horovodrun by initializing the processing group given the backend argument only in a general way.

With torch.multiprocessing.spawn

In this case idist Parallel is using the native torch torch.multiprocessing.spawn method under the hood in order to run the distributed configuration. Here nproc_per_node is passed as a spawn argument.

• Running multiple distributed configurations with one code. Source: ignite_idist.py:

# Running with gloo
python -u ignite_idist.py --nproc_per_node 2 --backend gloo

# Running with nccl
python -u ignite_idist.py --nproc_per_node 2 --backend nccl

# Running with horovod with gloo controller ( gloo or nccl support )
python -u ignite_idist.py --backend horovod --nproc_per_node 2

# Running on xla/tpu
python -u ignite_idist.py --backend xla-tpu --nproc_per_node 8 --batch_size 32

With Distributed launchers

PyTorch-Ignite’s idist Parallel context manager is also compatible with multiple distributed launchers.

With torchrun

Here we are using the torchrun script in order to spawn the processes:

torchrun --nproc_per_node 2 ignite_idist.py --backend gloo

With horovodrun

horovodrun -np 4 -H hostname1:2,hostname2:2 python ignite_idist.py --backend horovod

docker run --gpus all -it -v $PWD:/project pytorchignite/hvd-vision:latest /bin/bash
cd project

With slurm

The same result can be achieved by using slurm without any modification to the code:

srun --nodes=2
    --ntasks-per-node=2
    --job-name=pytorch-ignite
    --time=00:01:00
    --partition=gpgpu
    --gres=gpu:2
    --mem=10G
    python ignite_idist.py --backend nccl

or using sbatch script.bash with the script file script.bash:

#!/bin/bash
#SBATCH --job-name=pytorch-ignite
#SBATCH --output=slurm_%j.out
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=2
#SBATCH --time=00:01:00
#SBATCH --partition=gpgpu
#SBATCH --gres=gpu:2
#SBATCH --mem=10G

srun python ignite_idist.py --backend nccl

Closing Remarks

As we saw through the above examples, managing multiple configurations and specifications for distributed computing has never been easier. In just a few lines we can parallelize and execute code wherever it is while maintaining control and simplicity.

References

• idist-snippets: complete code used in this post.

• why-ignite: examples with distributed data parallel: native pytorch, pytorch-ignite, slurm.

• CIFAR10 example of distributed training on CIFAR10 with muliple configurations: 1 or multiple GPUs, multiple nodes and GPUs, TPUs.