🤗 Accelerate was created for PyTorch users who like to write the training loop of PyTorch models but are reluctant to write and maintain the boilerplate code needed to use multi-GPUs/TPU/fp16.
🤗 Accelerate abstracts exactly and only the boilerplate code related to multi-GPUs/TPU/fp16 and leaves the rest of your code unchanged.
Here is an example:
import torch
import torch.nn.functional as F
from datasets import load_dataset
+ from accelerate import Accelerator
+ accelerator = Accelerator()
- device = 'cpu'
+ device = accelerator.device
model = torch.nn.Transformer().to(device)
optimizer = torch.optim.Adam(model.parameters())
dataset = load_dataset('my_dataset')
data = torch.utils.data.DataLoader(dataset, shuffle=True)
+ model, optimizer, data = accelerator.prepare(model, optimizer, data)
model.train()
for epoch in range(10):
for source, targets in data:
source = source.to(device)
targets = targets.to(device)
optimizer.zero_grad()
output = model(source)
loss = F.cross_entropy(output, targets)
- loss.backward()
+ accelerator.backward(loss)
optimizer.step()As you can see in this example, by adding 5-lines to any standard PyTorch training script you can now run on any kind of single or distributed node setting (single CPU, single GPU, multi-GPUs and TPUs) as well as with or without mixed precision (fp16).
In particular, the same code can then be run without modification on your local machine for debugging or your training environment.
🤗 Accelerate even handles the device placement for you (which requires a few more changes to your code, but is safer in general), so you can even simplify your training loop further:
import torch
import torch.nn.functional as F
from datasets import load_dataset
+ from accelerate import Accelerator
- device = 'cpu'
+ accelerator = Accelerator()
- model = torch.nn.Transformer().to(device)
+ model = torch.nn.Transformer()
optimizer = torch.optim.Adam(model.parameters())
dataset = load_dataset('my_dataset')
data = torch.utils.data.DataLoader(dataset, shuffle=True)
+ model, optimizer, data = accelerator.prepare(model, optimizer, data)
model.train()
for epoch in range(10):
for source, targets in data:
- source = source.to(device)
- targets = targets.to(device)
optimizer.zero_grad()
output = model(source)
loss = F.cross_entropy(output, targets)
- loss.backward()
+ accelerator.backward(loss)
optimizer.step()Want to learn more? Check out the documentation or have look at our examples.
🤗 Accelerate also provides an optional CLI tool that allows you to quickly configure and test your training environment before launching the scripts. No need to remember how to use torch.distributed.launch or to write a specific launcher for TPU training!
On your machine(s) just run:
accelerate configand answer the questions asked. This will generate a config file that will be used automatically to properly set the default options when doing
accelerate launch my_script.py --args_to_my_scriptFor instance, here is how you would run the GLUE example on the MRPC task (from the root of the repo):
accelerate launch examples/nlp_example.pyThis CLI tool is optional, and you can still use python my_script.py or python -m torch.distributed.launch my_script.py at your convenance.
🤗 Here is another way to launch multi-CPU run using MPI. You can learn how to install Open MPI on this page. You can use Intel MPI or MVAPICH as well. Once you have MPI setup on your cluster, just run:
mpirun -np 2 python examples/nlp_example.py🤗 Accelerate supports training on single/multiple GPUs using DeepSpeed. To use it, you don't need to change anything in your training code; you can set everything using just accelerate config. However, if you desire to tweak your DeepSpeed related args from your python script, we provide you the DeepSpeedPlugin.
from accelerator import Accelerator, DeepSpeedPlugin
# deepspeed needs to know your gradient accumulation steps before hand, so don't forget to pass it
# Remember you still need to do gradient accumulation by yourself, just like you would have done without deepspeed
deepspeed_plugin = DeepSpeedPlugin(zero_stage=2, gradient_accumulation_steps=2)
accelerator = Accelerator(fp16=True, deepspeed_plugin=deepspeed_plugin)
# How to save your 🤗 Transformer?
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(save_dir, save_function=accelerator.save, state_dict=accelerator.get_state_dict(model))Note: DeepSpeed support is experimental for now. In case you get into some problem, please open an issue.
🤗 Accelerate also provides a notebook_launcher function you can use in a notebook to launch a distributed training. This is especially useful for Colab or Kaggle notebooks with a TPU backend. Just define your training loop in a training_function then in your last cell, add:
from accelerate import notebook_launcher
notebook_launcher(training_function)An example can be found in this notebook.
You should use 🤗 Accelerate when you want to easily run your training scripts in a distributed environment without having to renounce full control over your training loop. This is not a high-level framework above PyTorch, just a thin wrapper so you don't have to learn a new library, In fact the whole API of 🤗 Accelerate is in one class, the Accelerator object.
You shouldn't use 🤗 Accelerate if you don't want to write a training loop yourself. There are plenty of high-level libraries above PyTorch that will offer you that, 🤗 Accelerate is not one of them.
If you like the simplicity of 🤗 Accelerate but would prefer a higher-level abstraction around your training loop, some frameworks that are built on top of 🤗 Accelerate are listed below:
- Animus is a minimalistic framework to run machine learning experiments. Animus highlights common "breakpoints" in ML experiments and provides a unified interface for them within IExperiment.
- Catalyst is a PyTorch framework for Deep Learning Research and Development. It focuses on reproducibility, rapid experimentation, and codebase reuse so you can create something new rather than write yet another train loop. Catalyst provides a Runner to connect all parts of the experiment: hardware backend, data transformations, model train, and inference logic.
- fastai is a PyTorch framework for Deep Learning that simplifies training fast and accurate neural nets using modern best practices. fastai provides a Learner to handle the training, fine-tuning, and inference of deep learning algorithms.
- Kornia is a differentiable library that allows classical computer vision to be integrated into deep learning models. Kornia provides a Trainer with the specific purpose to train and fine-tune the supported deep learning algorithms within the library.
- pytorch-accelerated is a lightweight training library, with a streamlined feature set centred around a general-purpose Trainer, that places a huge emphasis on simplicity and transparency; enabling users to understand exactly what is going on under the hood, but without having to write and maintain the boilerplate themselves!
This repository is tested on Python 3.6+ and PyTorch 1.4.0+
You should install 🤗 Accelerate in a virtual environment. If you're unfamiliar with Python virtual environments, check out the user guide.
First, create a virtual environment with the version of Python you're going to use and activate it.
Then, you will need to install PyTorch: refer to the official installation page regarding the specific install command for your platform. Then 🤗 Accelerate can be installed using pip as follows:
pip install accelerate- CPU only
- multi-CPU on one node (machine)
- multi-CPU on several nodes (machines)
- single GPU
- multi-GPU on one node (machine)
- multi-GPU on several nodes (machines)
- TPU
- FP16 with native AMP (apex on the roadmap)
- DeepSpeed support (Experimental)
- PyTorch Fully Sharded Data Parallel (FSDP) support (Experimental)
- Megatron-LM support (Experimental)
If you use 🤗 Accelerate in your publication, please cite it by using the following BibTeX entry.
@Misc{accelerate,
title = {Accelerate: Training and inference at scale made simple, efficient and adaptable.},
author = {Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp Schmid, Zachary Mueller, Sourab Mangrulkar},
howpublished = {\url{https://github.com/huggingface/accelerate}},
year = {2022}
}