Let’s have a look at the 🤗 Accelerate main features and traps to avoid.
To use 🤗 Accelerate in your own script, you have to change four things:
- Import the Accelerator main class and instantiate one in an
from accelerate import Accelerator accelerator = Accelerator()
This should happen as early as possible in 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 (just one machine with a GPU, one machines with several GPUs, several machines with multiple GPUs or a TPU), the library will detect this automatically.
- Remove the call
.cuda()for your model and input data. The
acceleratorobject will handle this for you and place all those objects on the right device for you. If you know what you’re doing, you can leave those
.to(device)calls but you should use the device provided by the
To fully deactivate the automatic device placement, pass along
device_placement=False when initializing your
If you place your objects manually on the proper device, be careful to create your optimizer after putting your
accelerator.device or your training will fail on TPU.
- Pass all objects relevant to training (optimizer, model, training dataloader, learning rate scheduler) to the prepare() method. This will make sure everything is ready for training.
model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( model, optimizer, train_dataloader, lr_scheduler )
In particular, your training dataloader will be sharded across all GPUs/TPU cores available so that each one sees a
different portion of the training dataset. 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
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.
Alternatively, you can use the option
split_batches=True when creating and initializing your
Accelerator, in which case the batch size will always stay the same, whether you run your
script on 1, 2, 4, or 64 GPUs.
You should execute this instruction as soon as all objects for training are created, before starting your actual training loop.
You should only pass the learning rate scheduler to prepare() when the scheduler needs to be stepped at each optimizer step.
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
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().
You can perfectly send your dataloader to prepare() on its own, but it’s best to send the model and optimizer to prepare() together.
You may or may not want to send your validation dataloader to prepare(), depending on whether you want to run distributed evaluation or not (see below).
- Replace the line
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.
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
To perform distributed evaluation, send along your validation dataloader to the prepare() method:
validation_dataloader = accelerator.prepare(validation_dataloader)
As for your training dataloader, it will mean that (should you run your script on multiple devices) each device will only see part of the evaluation data. This means you will need to group your predictions together. This is very easy to 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
Any instruction using your training dataloader length (for instance if you need the number of total training steps to create a learning rate scheduler) should go after the call to prepare().
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.
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.
Launching your distributed script
You can use the regular commands to launch your distributed training (like
PyTorch), they are fully compatible with 🤗 Accelerate.
🤗 Accelerate also provides a CLI tool that unifies all launchers, so you only have to remember one command. To use it, just run:
on your machine and reply to the questions asked. This will save a default_config.yaml file in your cache folder for 🤗 Accelerate. That cache folder is (with decreasing order of priority):
- The content of your environment variable
HF_HOMEsuffixed with accelerate.
- If it does not exist, the content of your environment variable
XDG_CACHE_HOMEsuffixed with huggingface/accelerate.
- If this does not exist either, the folder ~/.cache/huggingface/accelerate
You can also specify with the flag
--config_file the location of the file you want to save.
Once this is done, you can test everything is going well on your setup by running:
This will launch a short script that will test the distributed environment. If it runs fine, 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 path_to_script.py --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 path_to_script.py --args_for_the_script
You can also 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.
Check out the Launch tutorial for more information about launching your scripts.
Launching training from a notebook
In Accelerate 0.3.0, a new notebook_launcher() has been introduced 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 (if the machine on which you are running your notebook has them).
Just 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 notebook_launcher(training_function)
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.
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
from accelerate import DistributedType if accelerator.distributed_type == DistributedType.TPU: # do something of static shape else: # 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.
We list here all smaller issues you could have in your script conversion and how to resolve them.
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 tqdm.auto 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
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 being sure every process is done with training. To do this, just write 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).
Saving/loading a model
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.
This is why it’s recommended to unwrap your model first. Here is an example:
accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) accelerator.save(unwrapped_model.state_dict(), filename)
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) unwrapped_model.load_state_dict(torch.load(filename))
Note that since all the model parameters are references to tensors, this will load your weights inside
Saving/loading entire states
When training your model, you may want to save the current state of the model, optimizer, random generators, and potentially LR 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
automatic_checkpoint_naming is enabled each saved checkpoint will be located then at
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
state_dict function to be stored
If you are using gradient clipping in your script, you should replace the calls to
torch.nn.utils.clip_grad_value_ with clipgrad_norm()
and clipgrad_value() respectively.
Mixed Precision training
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 accelerator.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: lr_scheduler.step()
To perform gradient accumulation use accumulate() and specify a
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) accelerator.backward(loss) optimizer.step() scheduler.step() optimizer.zero_grad()
DeepSpeed support is experimental, so the underlying API will evolve in the near future and may have some slight breaking changes. In particular, 🤗 Accelerate does not support DeepSpeed config you have written yourself yet, this will be added in a next version.
The notebook_launcher() does not support the DeepSpeed integration yet.
Internally, the library works by first analyzing the environment in which the script is launched to determine which
kind of distributed setup is used, how many different processes there are and which one the current script is in. All
that information is stored in the
This class is initialized the first time you instantiate an ~Accelerator as well as performing any specific initialization your distributed setup needs. Its state is then uniquely shared through all instances of AcceleratorState.
Then, when calling prepare(), the library:
- wraps your model(s) in the container adapted for the distributed setup,
- wraps your optimizer(s) in a AcceleratedOptimizer,
- creates a new version of your dataloader(s) in a DataLoaderShard.
While the model(s) and optimizer(s) are just put in simple wrappers, the dataloader(s) are re-created. This is mostly
because PyTorch does not let the user change the
batch_sampler of a dataloader once it’s been created and the
library handles the sharding of your data between processes by changing that
batch_sampler to yield every other
The DataLoaderShard subclasses
DataLoader to add the following functionality:
- it synchronizes the appropriate random number generator of all processes at each new iteration, to ensure any randomization (like shuffling) is done the exact same way across processes.
- it puts the batches on the proper device before yielding them (unless you have opted out of
The random number generator synchronization will by default synchronize:
generatorattribute of a given sampler (like the PyTorch
RandomSampler) for PyTorch >= 1.6
- the main random number generator in PyTorch <=1.5.1
You can choose which random number generator(s) to synchronize with the
rng_types argument of the main
Accelerator. In PyTorch >= 1.6, it is recommended to rely on a local
generator to avoid
setting the same seed in the main random number generator in all processes.
Synchronization of the main torch (or CUDA or XLA) random number generator will affect any other potential random artifacts you could have in your dataset (like random data augmentation) in the sense that all processes will get the same random numbers from the torch random modules (so will apply the same random data augmentation if it’s controlled by torch).
The randomization part of your custom sampler, batch sampler or iterable dataset should be done using a local
torch.Generator object (in PyTorch >= 1.6), see the traditional
RandomSampler, as an example.
For more details about the internals, see the Internals page.