Accelerate documentation

Quick tour

You are viewing v0.27.2 version. A newer version v0.30.1 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Quick tour

This guide aims to help you get started with 🤗 Accelerate quickly. It covers the essential steps you need to take to enable distributed training, as well as the adjustments that you need to make in some common scenarios.

To help you navigate, the guide is split into two sections:

Getting started with 🤗 Accelerate

Enable distributed training in your script

To use 🤗 Accelerate in your own training script, you have to modify four things:

  1. Import the Accelerator main class and instantiate one in an accelerator object.
from accelerate import Accelerator

accelerator = Accelerator()

Add this at the beginning of your training script as it will initialize everything necessary for distributed training. You don’t need to indicate the kind of environment you are in (a single machine with a GPU, a machine with several GPUs, or several machines with multiple GPUs or a TPU), the library will detect this automatically.

  1. Remove the .to(device) or .cuda() calls for your model and input data.

The accelerator object will handle placing these objects on the right device for you. If you choose to leave those .to(device) calls, make sure to use the device provided by the accelerator object: accelerator.device.

You can fully deactivate the automatic device placement by passing along device_placement=False when initializing the Accelerator. However, if you place your objects manually on the proper device, be careful to create your optimizer after putting your model on accelerator.device or your training will fail on TPU.

  1. Pass all PyTorch objects relevant to training (optimizer, model, dataloader(s), learning rate scheduler) to the prepare() method as soon as these objects are created, before starting your actual training loop:
model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
    model, optimizer, train_dataloader, lr_scheduler

Important notes:

  • You should always pass the the learning rate scheduler to prepare(), however if the scheduler should not be stepped at each optimization step, pass step_with_optimizer=False to the Accelerator init.
  • While you can send your dataloader to prepare() on its own (and there are cases for doing so, such as distributed inference), it’s best to send it to prepare() together with the model and optimizer.
  • If you wish to run distributed evaluation, send your validation dataloader to prepare() as well. There are some nuances to distributed validation, check the Distributed evaluation section of the guide.
  • Any instruction using your training dataloader length (for instance if you want to log the number of total training steps) should go after the call to prepare().

Passing DataLoader objects to the prepare() method ensures that your dataloader will be sharded across all GPUs/TPU cores available so that each one sees a different portion of the training dataset. In other words, if there are 8 processes and a dataset of 64 items, each process will see 8 of these items per iteration. Also, the random states of all processes will be synchronized at the beginning of each iteration through your dataloader, to make sure the data is shuffled the same way (if you decided to use shuffle=True or any kind of random sampler).

The actual batch size for your training will be the number of devices used multiplied by the batch size you set in your script. For instance, training on 4 GPUs with a batch size of 16 set when creating the training dataloader will train at an actual batch size of 64 (4 * 16). If you want the batch size remain the same regardless of how many GPUs the script is run on, you can use the option split_batches=True when creating and initializing Accelerator. Your training dataloader may change length when going through this method: if you run on X GPUs, it will have its length divided by X (since your actual batch size will be multiplied by X), unless you set split_batches=True.

  1. Replace the loss.backward() line with accelerator.backward(loss).

And you’re all set! With all these changes, your script will run on your local machine as well as on multiple GPUs or a TPU! You can either use your favorite tool to launch the distributed training, or you can use the 🤗 Accelerate launcher.

Add distributed evaluation

You can perform regular evaluation in your training script if you leave your validation dataloader out of the prepare() method. In this case, you will need to put the input data on the accelerator.device manually.

To perform distributed evaluation, send along your validation dataloader to the prepare() method:

validation_dataloader = accelerator.prepare(validation_dataloader)

Same as with your training dataloader, each device will only see part of the evaluation data should you run your script on multiple devices. This means you will need to group your predictions together which you can do with the gather_for_metrics() method.

for inputs, targets in validation_dataloader:
    predictions = model(inputs)
    # Gather all predictions and targets
    all_predictions, all_targets = accelerator.gather_for_metrics((predictions, targets))
    # Example of use with a *Datasets.Metric*
    metric.add_batch(all_predictions, all_targets)

Similar to the training dataloader, passing your validation dataloader through prepare() may change it: if you run on X GPUs, it will have its length divided by X (since your actual batch size will be multiplied by X), unless you set split_batches=True.

Some data at the end of the dataset may be duplicated so the batch can be divided equally among all workers. As a result, metrics should be calculated through the gather_for_metrics() method to automatically remove the duplicated data while gathering and provide a more accurate metric.

If for some reason you don’t wish to have this automatically done, gather() can be used instead to gather the data across all processes and this can manually be done instead.

The gather() and gather_for_metrics() methods require the tensors to be all the same size on each process. If you have tensors of different sizes on each process (for instance when dynamically padding to the maximum length in a batch), you should use the pad_across_processes() method to pad you tensor to the biggest size across processes.

Launch your distributed script

You can use the regular commands to launch your distributed training (like for PyTorch) - they are fully compatible with 🤗 Accelerate.

Alternatively, 🤗 Accelerate provides a CLI tool that unifies all launchers, so you only have to remember one command. \ To use it, run a quick configuration setup first on your machine and answer the questions:

accelerate config

At the end of the setup, a default_config.yaml file will be saved in your cache folder for 🤗 Accelerate. That cache folder is (with decreasing order of priority):

  • The content of your environment variable HF_HOME suffixed with accelerate.
  • If it does not exist, the content of your environment variable XDG_CACHE_HOME suffixed with huggingface/accelerate.
  • If this does not exist either, the folder ~/.cache/huggingface/accelerate.

By specifying the --config_file flag you can specify an alternative location of the configuration file. Once the configuration setup is complete, you can test your setup by running:

accelerate test

This will launch a short script that will test the distributed environment. If it runs without issues, you are ready for the next step!

Note that if you specified a location for the config file in the previous step, you need to pass it here as well:

accelerate test --config_file path_to_config.yaml

Now that this is done, you can run your script with the following command:

accelerate launch --args_for_the_script

If you stored the config file in a non-default location, you can indicate it to the launcher like this:

accelerate launch --config_file path_to_config.yaml --args_for_the_script

You can override any of the arguments determined by your config file. To see the complete list of parameters that you can pass in, run accelerate launch -h. (And further niche argument help by passing in partial commands, such as accelerate launch --multi_gpu -h for all multi_gpu args)

Check out the Launch tutorial for more information about launching your scripts.

Common modifications of the base case

The previous section covers the minimal essential steps to move a training script into a distributed setup with 🤗 Accelerate. Here we describe common modifications/deviations from the base case scenario and the adjustments you need to make to accommodate for them.

Launch distributed training from a notebook

Accelerate has a notebook_launcher() to help you launch your training function from a notebook. This launcher supports launching a training with TPUs on Colab or Kaggle, as well as training on several GPUs and machines (if the machine on which you are running your notebook has them).

Define a function responsible for your whole training and/or evaluation in a cell of the notebook, then execute a cell with the following code:

from accelerate import notebook_launcher


Your Accelerator object should only be defined inside the training function. This is because the initialization should be done inside the launcher only.

Check out the Notebook Launcher tutorial for more information about training on TPUs.

Specifics of training on TPU

If you want to launch your script on TPUs, there are a few caveats you should be aware of. Behind the scenes, the TPUs will create a graph of all the operations happening in your training step (forward pass, backward pass and optimizer step). This is why your first step of training will always be very long as building and compiling this graph for optimizations takes some time.

The good news is that this compilation will be cached so the second step and all the following will be much faster. The bad news is that it only applies if all of your steps do exactly the same operations, which implies:

  • having all tensors of the same length in all your batches
  • having static code (i.e., not a for loop of length that could change from step to step)

Having any of the things above change between two steps will trigger a new compilation which will, once again, take a lot of time. In practice, that means you must take special care to have all your tensors in your inputs of the same shape (so no dynamic padding for instance if you are in an NLP problem) and should not use layers with for loops that have different lengths depending on the inputs (such as an LSTM) or the training will be excruciatingly slow.

To introduce special behavior in your script for TPUs you can check the distributed_type of your accelerator:

from accelerate import DistributedType

if accelerator.distributed_type == DistributedType.TPU:
    # do something of static shape
    # go crazy and be dynamic

The NLP example shows an example in a situation with dynamic padding.

One last thing to pay close attention to: if your model has tied weights (such as language models which tie the weights of the embedding matrix with the weights of the decoder), moving this model to the TPU (either yourself or after you passed your model to prepare()) will break the tying. You will need to retie the weights after. You can find an example of this in the run_clm_no_trainer script in the Transformers repository.

Check out the TPU tutorial for more information about training on TPUs.

Execute a statement only on one processes

Some of your instructions only need to run for one process on a given server: for instance a data download or a log statement. To do this, wrap the statement in a test like this:

if accelerator.is_local_main_process:
    # Is executed once per server

Another example is progress bars: to avoid having multiple progress bars in your output, you should only display one on the local main process:

from import tqdm

progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)

The local means per machine: if you are running your training on two servers with several GPUs, the instruction will be executed once on each of those servers. If you need to execute something only once for all processes (and not per machine) for instance, uploading the final model to the 🤗 model hub, wrap it in a test like this:

if accelerator.is_main_process:
    # Is executed once only

For printing statements you only want executed once per machine, you can just replace the print function by accelerator.print.

Defer execution on multiple GPUs

When you run your usual script, instructions are executed in order. Using 🤗 Accelerate to deploy your script on several GPUs at the same time introduces a complication: while each process executes all instructions in order, some may be faster than others.

You might need to wait for all processes to have reached a certain point before executing a given instruction. For instance, you shouldn’t save a model before making sure every process is done with training. To do this, add the following line in your code:


This instruction will block all the processes that arrive first until all the other processes have reached that point (if you run your script on just one GPU or CPU, this won’t do anything).

Save/load a model in a distributed setup

Saving the model you trained might need a bit of adjustment: first you should wait for all processes to reach that point in the script as shown above, and then, you should unwrap your model before saving it. This is because when going through the prepare() method, your model may have been placed inside a bigger model, which deals with the distributed training. This in turn means that saving your model state dictionary without taking any precaution will take that potential extra layer into account, and you will end up with weights you can’t load back in your base model. The save_model() method will help you to achieve that. It will unwrap your model and save the model state dictionary.

Here is an example:

accelerator.save_model(model, save_directory)

The save_model() method can also save a model into sharded checkpoints or with safetensors format:

accelerator.save_model(model, save_directory, max_shard_size="1GB", safe_serialization=True)

If your script contains logic to load a checkpoint, we also recommend you load your weights in the unwrapped model (this is only useful if you use the load function after making your model go through prepare()). Here is an example:

unwrapped_model = accelerator.unwrap_model(model)
path_to_checkpoint = os.path.join(save_directory,"pytorch_model.bin")

Note that since all the model parameters are references to tensors, this will load your weights inside model.

If you want to load a sharded checkpoint or a checkpoint with safetensors format into the model with a specific device, we recommend you to load it with load_checkpoint_in_model() function. Here’s an example:

load_checkpoint_in_model(unwrapped_model, save_directory, device_map={"":device})

Save/load entire states

When training your model, you may want to save the current state of the model, optimizer, random generators, and potentially learning rate schedulers to be restored in the same script. You can use save_state() and load_state() respectively to do so.

To further customize where and how states saved through save_state() the ProjectConfiguration class can be used. For example if automatic_checkpoint_naming is enabled each saved checkpoint will be located then at Accelerator.project_dir/checkpoints/checkpoint_{checkpoint_number}.

If you have registered any other stateful items to be stored through register_for_checkpointing() they will also be saved and/or loaded.

Every object passed to register_for_checkpointing() must have a load_state_dict and state_dict function to be stored

Use gradient clipping

If you are using gradient clipping in your script, you should replace the calls to torch.nn.utils.clip_grad_norm_ or torch.nn.utils.clip_grad_value_ with clipgrad_norm() and clipgrad_value() respectively.

Train with mixed precision

If you are running your training in Mixed Precision with 🤗 Accelerate, you will get the best result with your loss being computed inside your model (like in Transformer models for instance). Every computation outside of the model will be executed in full precision (which is generally what you want for loss computation, especially if it involves a softmax). However, you might want to put your loss computation inside the autocast() context manager:

with accelerator.autocast():
    loss = complex_loss_function(outputs, target):

Another caveat with Mixed Precision training is that the gradient will skip a few updates at the beginning and sometimes during training: because of the dynamic loss scaling strategy, there are points during training where the gradients have overflown, and the loss scaling factor is reduced to avoid this happening again at the next step.

This means that you may update your learning rate scheduler when there was no update, which is fine in general, but may have an impact when you have very little training data, or if the first learning rate values of your scheduler are very important. In this case, you can skip the learning rate scheduler updates when the optimizer step was not done like this:

if not accelerator.optimizer_step_was_skipped:

Use gradient accumulation

To perform gradient accumulation use accumulate() and specify a gradient_accumulation_steps. This will also automatically ensure the gradients are synced or unsynced when on multi-device training, check if the step should actually be performed, and auto-scale the loss:

accelerator = Accelerator(gradient_accumulation_steps=2)
model, optimizer, training_dataloader = accelerator.prepare(model, optimizer, training_dataloader)

for input, label in training_dataloader:
    with accelerator.accumulate(model):
        predictions = model(input)
        loss = loss_function(predictions, label)