Training on TPUs with 🤗 Accelerate

Training on TPUs can be slightly different than training on multi-gpu, even with 🤗 Accelerate. This guide aims to show you where you should be careful and why, as well as the best practices in general.

Training in a Notebook

The main carepoint when training on TPUs comes from the notebook_launcher(). As mentioned in the notebook tutorial, you need to restructure your training code into a function that can get passed to the notebook_launcher() function and be careful about not declaring any tensors on the GPU.

While on a TPU that last part is not as important, a critical part to understand is that when you launch code from a notebook you do so through a process called forking. When launching from the command-line, you perform spawning, where a python process is not currently running and you spawn a new process in. Since your Jupyter notebook is already utilizing a python process, you need to fork a new process from it to launch your code.

Where this becomes important is in regards to declaring your model. On forked TPU processes, it is recommended that you instantiate your model once and pass this into your training function. This is different than training on GPUs where you create n models that have their gradients synced and back-propagated at certain moments. Instead one model instance is shared between all the nodes and it is passed back and forth. This is important especially when training on low-resource TPUs such as those provided in Kaggle kernels or on Google Colaboratory.

Below is an example of a training function passed to the notebook_launcher() if training on CPUs or GPUs:

def training_function():
    # Initialize accelerator
    accelerator = Accelerator()
    model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=2)
    train_dataloader, eval_dataloader = create_dataloaders(
        train_batch_size=hyperparameters["train_batch_size"], eval_batch_size=hyperparameters["eval_batch_size"]
    )

    # Instantiate optimizer
    optimizer = AdamW(params=model.parameters(), lr=hyperparameters["learning_rate"])

    # Prepare everything
    # There is no specific order to remember, we just need to unpack the objects in the same order we gave them to the
    # prepare method.
    model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader
    )

    num_epochs = hyperparameters["num_epochs"]
    # Now we train the model
    for epoch in range(num_epochs):
        model.train()
        for step, batch in enumerate(train_dataloader):
            outputs = model(**batch)
            loss = outputs.loss
            accelerator.backward(loss)

            optimizer.step()
            optimizer.zero_grad()

from accelerate import notebook_launcher

notebook_launcher(training_function)

If you use this example and declare the model inside the training loop, then on a low-resource system you will potentially see an error like:

ProcessExitedException: process 0 terminated with signal SIGSEGV

This error is extremely cryptic but the basic explanation is you ran out of system RAM. You can avoid this entirely by reconfiguring the training function to accept a single model argument, and declare it in an outside cell:

# In another Jupyter cell
model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=2)

+ def training_function(model):
      # Initialize accelerator
      accelerator = Accelerator()
-     model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=2)
      train_dataloader, eval_dataloader = create_dataloaders(
          train_batch_size=hyperparameters["train_batch_size"], eval_batch_size=hyperparameters["eval_batch_size"]
      )
  ...

And finally calling the training function with:

  from accelerate import notebook_launcher
- notebook_launcher(training_function)
+ notebook_launcher(training_function, (model,))

Mixed Precision and Global Variables

As mentioned in the mixed precision tutorial, 🤗 Accelerate supports fp16 and bf16, both of which can be used on TPUs. That being said, ideally bf16 should be utilized as it is extremely efficient to use.

There are two “layers” when using bf16 and 🤗 Accelerate on TPUs, at the base level and at the operation level.

At the base level, this is enabled when passing mixed_precision="bf16" to Accelerator, such as:

accelerator = Accelerator(mixed_precision="bf16")

By default this will cast torch.float and torch.double to bfloat16 on TPUs. The specific configuration being set is an environmental variable of XLA_USE_BF16 is set to 1.

There is a further configuration you can perform which is setting the XLA_DOWNCAST_BF16 environmental variable. If set to 1, then torch.float is bfloat16 and torch.double is float32.

This is performed in the Accelerator object when passing downcast_bf16=True:

accelerator = Accelerator(mixed_precision="bf16", downcast_bf16=True)

Using downcasting instead of bf16 everywhere is good for when you are trying to calculate metrics, log values, and more where raw bf16 tensors would be unusable.

Training Times on TPUs

As you launch your script, you may notice that training seems exceptionally slow at first. This is because TPUs first run through a few batches of data to see how much memory to allocate before finally utilizing this configured memory allocation extremely efficiently.

If you notice that your evaluation code to calculate the metrics of your model takes longer due to a larger batch size being used, it is recommended to keep the batch size the same as the training data if it is too slow. Otherwise the memory will reallocate to this new batch size after the first few iterations.