Transformers documentation

Trainer

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Trainer

The Trainer class provides an API for feature-complete training in PyTorch for most standard use cases. It’s used in most of the example scripts.

Before instantiating your Trainer, create a TrainingArguments to access all the points of customization during training.

The API supports distributed training on multiple GPUs/TPUs, mixed precision through NVIDIA Apex and Native AMP for PyTorch.

The Trainer contains the basic training loop which supports the above features. To inject custom behavior you can subclass them and override the following methods:

  • get_train_dataloader — Creates the training DataLoader.
  • get_eval_dataloader — Creates the evaluation DataLoader.
  • get_test_dataloader — Creates the test DataLoader.
  • log — Logs information on the various objects watching training.
  • create_optimizer_and_scheduler — Sets up the optimizer and learning rate scheduler if they were not passed at init. Note, that you can also subclass or override the create_optimizer and create_scheduler methods separately.
  • create_optimizer — Sets up the optimizer if it wasn’t passed at init.
  • create_scheduler — Sets up the learning rate scheduler if it wasn’t passed at init.
  • compute_loss - Computes the loss on a batch of training inputs.
  • training_step — Performs a training step.
  • prediction_step — Performs an evaluation/test step.
  • evaluate — Runs an evaluation loop and returns metrics.
  • predict — Returns predictions (with metrics if labels are available) on a test set.

The Trainer class is optimized for 🤗 Transformers models and can have surprising behaviors when you use it on other models. When using it on your own model, make sure:

  • your model always return tuples or subclasses of ModelOutput.
  • your model can compute the loss if a labels argument is provided and that loss is returned as the first element of the tuple (if your model returns tuples)
  • your model can accept multiple label arguments (use the label_names in your TrainingArguments to indicate their name to the Trainer) but none of them should be named "label".

Here is an example of how to customize Trainer to use a weighted loss (useful when you have an unbalanced training set):

from torch import nn
from transformers import Trainer


class CustomTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False):
        labels = inputs.get("labels")
        # forward pass
        outputs = model(**inputs)
        logits = outputs.get("logits")
        # compute custom loss (suppose one has 3 labels with different weights)
        loss_fct = nn.CrossEntropyLoss(weight=torch.tensor([1.0, 2.0, 3.0]))
        loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
        return (loss, outputs) if return_outputs else loss

Another way to customize the training loop behavior for the PyTorch Trainer is to use callbacks that can inspect the training loop state (for progress reporting, logging on TensorBoard or other ML platforms…) and take decisions (like early stopping).

Trainer

class transformers.Trainer

< >

( model: typing.Union[transformers.modeling_utils.PreTrainedModel, torch.nn.modules.module.Module] = None args: TrainingArguments = None data_collator: typing.Optional[DataCollator] = None train_dataset: typing.Optional[torch.utils.data.dataset.Dataset] = None eval_dataset: typing.Optional[torch.utils.data.dataset.Dataset] = None tokenizer: typing.Optional[transformers.tokenization_utils_base.PreTrainedTokenizerBase] = None model_init: typing.Callable[[], transformers.modeling_utils.PreTrainedModel] = None compute_metrics: typing.Union[typing.Callable[[transformers.trainer_utils.EvalPrediction], typing.Dict], NoneType] = None callbacks: typing.Optional[typing.List[transformers.trainer_callback.TrainerCallback]] = None optimizers: typing.Tuple[torch.optim.optimizer.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None) preprocess_logits_for_metrics: typing.Callable[[torch.Tensor, torch.Tensor], torch.Tensor] = None )

Parameters

  • model (PreTrainedModel or torch.nn.Module, optional) — The model to train, evaluate or use for predictions. If not provided, a model_init must be passed.

    Trainer is optimized to work with the PreTrainedModel provided by the library. You can still use your own models defined as torch.nn.Module as long as they work the same way as the 🤗 Transformers models.

  • args (TrainingArguments, optional) — The arguments to tweak for training. Will default to a basic instance of TrainingArguments with the output_dir set to a directory named tmp_trainer in the current directory if not provided.
  • data_collator (DataCollator, optional) — The function to use to form a batch from a list of elements of train_dataset or eval_dataset. Will default to default_data_collator() if no tokenizer is provided, an instance of DataCollatorWithPadding otherwise.
  • train_dataset (torch.utils.data.Dataset or torch.utils.data.IterableDataset, optional) — The dataset to use for training. If it is a Dataset, columns not accepted by the model.forward() method are automatically removed.

    Note that if it’s a torch.utils.data.IterableDataset with some randomization and you are training in a distributed fashion, your iterable dataset should either use a internal attribute generator that is a torch.Generator for the randomization that must be identical on all processes (and the Trainer will manually set the seed of this generator at each epoch) or have a set_epoch() method that internally sets the seed of the RNGs used.

  • eval_dataset (torch.utils.data.Dataset, optional) — The dataset to use for evaluation. If it is a Dataset, columns not accepted by the model.forward() method are automatically removed.
  • tokenizer (PreTrainedTokenizerBase, optional) — The tokenizer used to preprocess the data. If provided, will be used to automatically pad the inputs the maximum length when batching inputs, and it will be saved along the model to make it easier to rerun an interrupted training or reuse the fine-tuned model.
  • model_init (Callable[[], PreTrainedModel], optional) — A function that instantiates the model to be used. If provided, each call to train() will start from a new instance of the model as given by this function.

    The function may have zero argument, or a single one containing the optuna/Ray Tune/SigOpt trial object, to be able to choose different architectures according to hyper parameters (such as layer count, sizes of inner layers, dropout probabilities etc).

  • compute_metrics (Callable[[EvalPrediction], Dict], optional) — The function that will be used to compute metrics at evaluation. Must take a EvalPrediction and return a dictionary string to metric values.
  • callbacks (List of TrainerCallback, optional) — A list of callbacks to customize the training loop. Will add those to the list of default callbacks detailed in here.

    If you want to remove one of the default callbacks used, use the Trainer.remove_callback() method.

  • optimizers (Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR], optional) — A tuple containing the optimizer and the scheduler to use. Will default to an instance of AdamW on your model and a scheduler given by get_linear_schedule_with_warmup() controlled by args.
  • preprocess_logits_for_metrics (Callable[[torch.Tensor, torch.Tensor], torch.Tensor], optional) — A function that preprocess the logits right before caching them at each evaluation step. Must take two tensors, the logits and the labels, and return the logits once processed as desired. The modifications made by this function will be reflected in the predictions received by compute_metrics.

    Note that the labels (second parameter) will be None if the dataset does not have them.

Trainer is a simple but feature-complete training and eval loop for PyTorch, optimized for 🤗 Transformers.

Important attributes:

  • model — Always points to the core model. If using a transformers model, it will be a PreTrainedModel subclass.
  • model_wrapped — Always points to the most external model in case one or more other modules wrap the original model. This is the model that should be used for the forward pass. For example, under DeepSpeed, the inner model is wrapped in DeepSpeed and then again in torch.nn.DistributedDataParallel. If the inner model hasn’t been wrapped, then self.model_wrapped is the same as self.model.
  • is_model_parallel — Whether or not a model has been switched to a model parallel mode (different from data parallelism, this means some of the model layers are split on different GPUs).
  • place_model_on_device — Whether or not to automatically place the model on the device - it will be set to False if model parallel or deepspeed is used, or if the default TrainingArguments.place_model_on_device is overridden to return False .
  • is_in_train — Whether or not a model is currently running train (e.g. when evaluate is called while in train)

add_callback

< >

( callback )

Parameters

  • callback (type or TrainerCallback) — A TrainerCallback class or an instance of a TrainerCallback. In the first case, will instantiate a member of that class.

Add a callback to the current list of TrainerCallback.

autocast_smart_context_manager

< >

( )

A helper wrapper that creates an appropriate context manager for autocast while feeding it the desired arguments, depending on the situation.

compute_loss

< >

( model inputs return_outputs = False )

How the loss is computed by Trainer. By default, all models return the loss in the first element.

Subclass and override for custom behavior.

compute_loss_context_manager

< >

( )

A helper wrapper to group together context managers.

create_optimizer

< >

( )

Setup the optimizer.

We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the Trainer’s init through optimizers, or subclass and override this method in a subclass.

create_optimizer_and_scheduler

< >

( num_training_steps: int )

Setup the optimizer and the learning rate scheduler.

We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the Trainer’s init through optimizers, or subclass and override this method (or create_optimizer and/or create_scheduler) in a subclass.

create_scheduler

< >

( num_training_steps: int optimizer: Optimizer = None )

Parameters

  • num_training_steps (int) — The number of training steps to do.

Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or passed as an argument.

evaluate

< >

( eval_dataset: typing.Optional[torch.utils.data.dataset.Dataset] = None ignore_keys: typing.Optional[typing.List[str]] = None metric_key_prefix: str = 'eval' )

Parameters

  • eval_dataset (Dataset, optional) — Pass a dataset if you wish to override self.eval_dataset. If it is a Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement the __len__ method.
  • ignore_keys (Lst[str], optional) — A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.
  • metric_key_prefix (str, optional, defaults to "eval") — An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is “eval” (default)

Run evaluation and returns metrics.

The calling script will be responsible for providing a method to compute metrics, as they are task-dependent (pass it to the init compute_metrics argument).

You can also subclass and override this method to inject custom behavior.

evaluation_loop

< >

( dataloader: DataLoader description: str prediction_loss_only: typing.Optional[bool] = None ignore_keys: typing.Optional[typing.List[str]] = None metric_key_prefix: str = 'eval' )

Prediction/evaluation loop, shared by Trainer.evaluate() and Trainer.predict().

Works both with or without labels.

floating_point_ops

< >

( inputs: typing.Dict[str, typing.Union[torch.Tensor, typing.Any]] ) int

Parameters

  • inputs (Dict[str, Union[torch.Tensor, Any]]) — The inputs and targets of the model.

Returns

int

The number of floating-point operations.

For models that inherit from PreTrainedModel, uses that method to compute the number of floating point operations for every backward + forward pass. If using another model, either implement such a method in the model or subclass and override this method.

get_eval_dataloader

< >

( eval_dataset: typing.Optional[torch.utils.data.dataset.Dataset] = None )

Parameters

  • eval_dataset (torch.utils.data.Dataset, optional) — If provided, will override self.eval_dataset. If it is a Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement __len__.

Returns the evaluation DataLoader.

Subclass and override this method if you want to inject some custom behavior.

get_optimizer_cls_and_kwargs

< >

( args: TrainingArguments )

Parameters

  • args (transformers.training_args.TrainingArguments) — The training arguments for the training session.

Returns the optimizer class and optimizer parameters based on the training arguments.

get_test_dataloader

< >

( test_dataset: Dataset )

Parameters

  • test_dataset (torch.utils.data.Dataset, optional) — The test dataset to use. If it is a Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement __len__.

Returns the test DataLoader.

Subclass and override this method if you want to inject some custom behavior.

get_train_dataloader

< >

( )

Returns the training DataLoader.

Will use no sampler if train_dataset does not implement __len__, a random sampler (adapted to distributed training if necessary) otherwise.

Subclass and override this method if you want to inject some custom behavior.

hyperparameter_search

< >

( hp_space: typing.Union[typing.Callable[[ForwardRef('optuna.Trial')], typing.Dict[str, float]], NoneType] = None compute_objective: typing.Union[typing.Callable[[typing.Dict[str, float]], float], NoneType] = None n_trials: int = 20 direction: str = 'minimize' backend: typing.Union[ForwardRef('str'), transformers.trainer_utils.HPSearchBackend, NoneType] = None hp_name: typing.Union[typing.Callable[[ForwardRef('optuna.Trial')], str], NoneType] = None **kwargs ) trainer_utils.BestRun

Parameters

  • hp_space (Callable[["optuna.Trial"], Dict[str, float]], optional) — A function that defines the hyperparameter search space. Will default to default_hp_space_optuna() or default_hp_space_ray() or default_hp_space_sigopt() depending on your backend.
  • compute_objective (Callable[[Dict[str, float]], float], optional) — A function computing the objective to minimize or maximize from the metrics returned by the evaluate method. Will default to default_compute_objective().
  • n_trials (int, optional, defaults to 100) — The number of trial runs to test.
  • direction(str, optional, defaults to "minimize") — Whether to optimize greater or lower objects. Can be "minimize" or "maximize", you should pick "minimize" when optimizing the validation loss, "maximize" when optimizing one or several metrics.
  • backend(str or HPSearchBackend, optional) — The backend to use for hyperparameter search. Will default to optuna or Ray Tune or SigOpt, depending on which one is installed. If all are installed, will default to optuna. kwargs — Additional keyword arguments passed along to optuna.create_study or ray.tune.run. For more information see:

Returns

trainer_utils.BestRun

All the information about the best run.

Launch an hyperparameter search using optuna or Ray Tune or SigOpt. The optimized quantity is determined by compute_objective, which defaults to a function returning the evaluation loss when no metric is provided, the sum of all metrics otherwise.

To use this method, you need to have provided a model_init when initializing your Trainer: we need to reinitialize the model at each new run. This is incompatible with the optimizers argument, so you need to subclass Trainer and override the method create_optimizer_and_scheduler() for custom optimizer/scheduler.

init_git_repo

< >

( at_init: bool = False )

Parameters

  • at_init (bool, optional, defaults to False) — Whether this function is called before any training or not. If self.args.overwrite_output_dir is True and at_init is True, the path to the repo (which is self.args.output_dir) might be wiped out.

Initializes a git repo in self.args.hub_model_id.

is_local_process_zero

< >

( )

Whether or not this process is the local (e.g., on one machine if training in a distributed fashion on several machines) main process.

is_world_process_zero

< >

( )

Whether or not this process is the global main process (when training in a distributed fashion on several machines, this is only going to be True for one process).

log

< >

( logs: typing.Dict[str, float] )

Parameters

  • logs (Dict[str, float]) — The values to log.

Log logs on the various objects watching training.

Subclass and override this method to inject custom behavior.

log_metrics

< >

( split metrics )

Parameters

  • split (str) — Mode/split name: one of train, eval, test
  • metrics (Dict[str, float]) — The metrics returned from train/evaluate/predictmetrics: metrics dict

Log metrics in a specially formatted way

Under distributed environment this is done only for a process with rank 0.

Notes on memory reports:

In order to get memory usage report you need to install psutil. You can do that with pip install psutil.

Now when this method is run, you will see a report that will include: :

init_mem_cpu_alloc_delta   =     1301MB
init_mem_cpu_peaked_delta  =      154MB
init_mem_gpu_alloc_delta   =      230MB
init_mem_gpu_peaked_delta  =        0MB
train_mem_cpu_alloc_delta  =     1345MB
train_mem_cpu_peaked_delta =        0MB
train_mem_gpu_alloc_delta  =      693MB
train_mem_gpu_peaked_delta =        7MB

Understanding the reports:

  • the first segment, e.g., train__, tells you which stage the metrics are for. Reports starting with init_ will be added to the first stage that gets run. So that if only evaluation is run, the memory usage for the __init__ will be reported along with the eval_ metrics.
  • the third segment, is either cpu or gpu, tells you whether it’s the general RAM or the gpu0 memory metric.
  • *_alloc_delta - is the difference in the used/allocated memory counter between the end and the start of the stage - it can be negative if a function released more memory than it allocated.
  • *_peaked_delta - is any extra memory that was consumed and then freed - relative to the current allocated memory counter - it is never negative. When you look at the metrics of any stage you add up alloc_delta + peaked_delta and you know how much memory was needed to complete that stage.

The reporting happens only for process of rank 0 and gpu 0 (if there is a gpu). Typically this is enough since the main process does the bulk of work, but it could be not quite so if model parallel is used and then other GPUs may use a different amount of gpu memory. This is also not the same under DataParallel where gpu0 may require much more memory than the rest since it stores the gradient and optimizer states for all participating GPUS. Perhaps in the future these reports will evolve to measure those too.

The CPU RAM metric measures RSS (Resident Set Size) includes both the memory which is unique to the process and the memory shared with other processes. It is important to note that it does not include swapped out memory, so the reports could be imprecise.

The CPU peak memory is measured using a sampling thread. Due to python’s GIL it may miss some of the peak memory if that thread didn’t get a chance to run when the highest memory was used. Therefore this report can be less than reality. Using tracemalloc would have reported the exact peak memory, but it doesn’t report memory allocations outside of python. So if some C++ CUDA extension allocated its own memory it won’t be reported. And therefore it was dropped in favor of the memory sampling approach, which reads the current process memory usage.

The GPU allocated and peak memory reporting is done with torch.cuda.memory_allocated() and torch.cuda.max_memory_allocated(). This metric reports only “deltas” for pytorch-specific allocations, as torch.cuda memory management system doesn’t track any memory allocated outside of pytorch. For example, the very first cuda call typically loads CUDA kernels, which may take from 0.5 to 2GB of GPU memory.

Note that this tracker doesn’t account for memory allocations outside of Trainer’s __init__, train, evaluate and predict calls.

Because evaluation calls may happen during train, we can’t handle nested invocations because torch.cuda.max_memory_allocated is a single counter, so if it gets reset by a nested eval call, train’s tracker will report incorrect info. If this pytorch issue gets resolved it will be possible to change this class to be re-entrant. Until then we will only track the outer level of train, evaluate and predict methods. Which means that if eval is called during train, it’s the latter that will account for its memory usage and that of the former.

This also means that if any other tool that is used along the Trainer calls torch.cuda.reset_peak_memory_stats, the gpu peak memory stats could be invalid. And the Trainer will disrupt the normal behavior of any such tools that rely on calling torch.cuda.reset_peak_memory_stats themselves.

For best performance you may want to consider turning the memory profiling off for production runs.

metrics_format

< >

( metrics: typing.Dict[str, float] ) metrics (Dict[str, float])

Parameters

  • metrics (Dict[str, float]) — The metrics returned from train/evaluate/predict

Returns

metrics (Dict[str, float])

The reformatted metrics

Reformat Trainer metrics values to a human-readable format

num_examples

< >

( dataloader: DataLoader )

Helper to get number of samples in a DataLoader by accessing its dataset. When dataloader.dataset does not exist or has no length, estimates as best it can

pop_callback

< >

( callback ) TrainerCallback

Parameters

  • callback (type or TrainerCallback) — A TrainerCallback class or an instance of a TrainerCallback. In the first case, will pop the first member of that class found in the list of callbacks.

Returns

TrainerCallback

The callback removed, if found.

Remove a callback from the current list of TrainerCallback and returns it.

If the callback is not found, returns None (and no error is raised).

predict

< >

( test_dataset: Dataset ignore_keys: typing.Optional[typing.List[str]] = None metric_key_prefix: str = 'test' )

Parameters

  • test_dataset (Dataset) — Dataset to run the predictions on. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed. Has to implement the method __len__
  • ignore_keys (Lst[str], optional) — A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.
  • metric_key_prefix (str, optional, defaults to "test") — An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “test_bleu” if the prefix is “test” (default)

Run prediction and returns predictions and potential metrics.

Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method will also return metrics, like in evaluate().

If your predictions or labels have different sequence length (for instance because you’re doing dynamic padding in a token classification task) the predictions will be padded (on the right) to allow for concatenation into one array. The padding index is -100.

Returns: NamedTuple A namedtuple with the following keys:

  • predictions (np.ndarray): The predictions on test_dataset.
  • label_ids (np.ndarray, optional): The labels (if the dataset contained some).
  • metrics (Dict[str, float], optional): The potential dictionary of metrics (if the dataset contained labels).

prediction_loop

< >

( dataloader: DataLoader description: str prediction_loss_only: typing.Optional[bool] = None ignore_keys: typing.Optional[typing.List[str]] = None metric_key_prefix: str = 'eval' )

Prediction/evaluation loop, shared by Trainer.evaluate() and Trainer.predict().

Works both with or without labels.

prediction_step

< >

( model: Module inputs: typing.Dict[str, typing.Union[torch.Tensor, typing.Any]] prediction_loss_only: bool ignore_keys: typing.Optional[typing.List[str]] = None ) Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]

Parameters

  • model (nn.Module) — The model to evaluate.
  • inputs (Dict[str, Union[torch.Tensor, Any]]) — The inputs and targets of the model.

    The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument labels. Check your model’s documentation for all accepted arguments.

  • prediction_loss_only (bool) — Whether or not to return the loss only.
  • ignore_keys (Lst[str], optional) — A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.

Returns

Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]

A tuple with the loss, logits and labels (each being optional).

Perform an evaluation step on model using inputs.

Subclass and override to inject custom behavior.

push_to_hub

< >

( commit_message: typing.Optional[str] = 'End of training' blocking: bool = True **kwargs )

Parameters

  • commit_message (str, optional, defaults to "End of training") — Message to commit while pushing.
  • blocking (bool, optional, defaults to True) — Whether the function should return only when the git push has finished. kwargs — Additional keyword arguments passed along to create_model_card().

Upload self.model and self.tokenizer to the 🤗 model hub on the repo self.args.hub_model_id.

remove_callback

< >

( callback )

Parameters

  • callback (type or TrainerCallback) — A TrainerCallback class or an instance of a TrainerCallback. In the first case, will remove the first member of that class found in the list of callbacks.

Remove a callback from the current list of TrainerCallback.

save_metrics

< >

( split metrics combined = True )

Parameters

  • split (str) — Mode/split name: one of train, eval, test, all
  • metrics (Dict[str, float]) — The metrics returned from train/evaluate/predict
  • combined (bool, optional, defaults to True) — Creates combined metrics by updating all_results.json with metrics of this call

Save metrics into a json file for that split, e.g. train_results.json.

Under distributed environment this is done only for a process with rank 0.

To understand the metrics please read the docstring of log_metrics(). The only difference is that raw unformatted numbers are saved in the current method.

save_model

< >

( output_dir: typing.Optional[str] = None _internal_call: bool = False )

Will save the model, so you can reload it using from_pretrained().

Will only save from the main process.

save_state

< >

( )

Saves the Trainer state, since Trainer.save_model saves only the tokenizer with the model

Under distributed environment this is done only for a process with rank 0.

torchdynamo_smart_context_manager

< >

( )

A helper wrapper that creates an appropriate context manager for torchdynamo.

train

< >

( resume_from_checkpoint: typing.Union[str, bool, NoneType] = None trial: typing.Union[ForwardRef('optuna.Trial'), typing.Dict[str, typing.Any]] = None ignore_keys_for_eval: typing.Optional[typing.List[str]] = None **kwargs )

Parameters

  • resume_from_checkpoint (str or bool, optional) — If a str, local path to a saved checkpoint as saved by a previous instance of Trainer. If a bool and equals True, load the last checkpoint in args.output_dir as saved by a previous instance of Trainer. If present, training will resume from the model/optimizer/scheduler states loaded here.
  • trial (optuna.Trial or Dict[str, Any], optional) — The trial run or the hyperparameter dictionary for hyperparameter search.
  • ignore_keys_for_eval (List[str], optional) — A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions for evaluation during the training. kwargs — Additional keyword arguments used to hide deprecated arguments

Main training entry point.

training_step

< >

( model: Module inputs: typing.Dict[str, typing.Union[torch.Tensor, typing.Any]] ) torch.Tensor

Parameters

  • model (nn.Module) — The model to train.
  • inputs (Dict[str, Union[torch.Tensor, Any]]) — The inputs and targets of the model.

    The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument labels. Check your model’s documentation for all accepted arguments.

Returns

torch.Tensor

The tensor with training loss on this batch.

Perform a training step on a batch of inputs.

Subclass and override to inject custom behavior.

Seq2SeqTrainer

class transformers.Seq2SeqTrainer

< >

( model: typing.Union[transformers.modeling_utils.PreTrainedModel, torch.nn.modules.module.Module] = None args: TrainingArguments = None data_collator: typing.Optional[DataCollator] = None train_dataset: typing.Optional[torch.utils.data.dataset.Dataset] = None eval_dataset: typing.Optional[torch.utils.data.dataset.Dataset] = None tokenizer: typing.Optional[transformers.tokenization_utils_base.PreTrainedTokenizerBase] = None model_init: typing.Callable[[], transformers.modeling_utils.PreTrainedModel] = None compute_metrics: typing.Union[typing.Callable[[transformers.trainer_utils.EvalPrediction], typing.Dict], NoneType] = None callbacks: typing.Optional[typing.List[transformers.trainer_callback.TrainerCallback]] = None optimizers: typing.Tuple[torch.optim.optimizer.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None) preprocess_logits_for_metrics: typing.Callable[[torch.Tensor, torch.Tensor], torch.Tensor] = None )

evaluate

< >

( eval_dataset: typing.Optional[torch.utils.data.dataset.Dataset] = None ignore_keys: typing.Optional[typing.List[str]] = None metric_key_prefix: str = 'eval' **gen_kwargs )

Parameters

  • eval_dataset (Dataset, optional) — Pass a dataset if you wish to override self.eval_dataset. If it is an Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement the __len__ method.
  • ignore_keys (List[str], optional) — A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.
  • metric_key_prefix (str, optional, defaults to "eval") — An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is "eval" (default)
  • max_length (int, optional) — The maximum target length to use when predicting with the generate method.
  • num_beams (int, optional) — Number of beams for beam search that will be used when predicting with the generate method. 1 means no beam search. gen_kwargs — Additional generate specific kwargs.

Run evaluation and returns metrics.

The calling script will be responsible for providing a method to compute metrics, as they are task-dependent (pass it to the init compute_metrics argument).

You can also subclass and override this method to inject custom behavior.

predict

< >

( test_dataset: Dataset ignore_keys: typing.Optional[typing.List[str]] = None metric_key_prefix: str = 'test' **gen_kwargs )

Parameters

  • test_dataset (Dataset) — Dataset to run the predictions on. If it is a Dataset, columns not accepted by the model.forward() method are automatically removed. Has to implement the method __len__
  • ignore_keys (List[str], optional) — A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.
  • metric_key_prefix (str, optional, defaults to "eval") — An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is "eval" (default)
  • max_length (int, optional) — The maximum target length to use when predicting with the generate method.
  • num_beams (int, optional) — Number of beams for beam search that will be used when predicting with the generate method. 1 means no beam search. gen_kwargs — Additional generate specific kwargs.

Run prediction and returns predictions and potential metrics.

Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method will also return metrics, like in evaluate().

If your predictions or labels have different sequence lengths (for instance because you’re doing dynamic padding in a token classification task) the predictions will be padded (on the right) to allow for concatenation into one array. The padding index is -100.

Returns: NamedTuple A namedtuple with the following keys:

  • predictions (np.ndarray): The predictions on test_dataset.
  • label_ids (np.ndarray, optional): The labels (if the dataset contained some).
  • metrics (Dict[str, float], optional): The potential dictionary of metrics (if the dataset contained labels).

TrainingArguments

class transformers.TrainingArguments

< >

( output_dir: str overwrite_output_dir: bool = False do_train: bool = False do_eval: bool = False do_predict: bool = False evaluation_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'no' prediction_loss_only: bool = False per_device_train_batch_size: int = 8 per_device_eval_batch_size: int = 8 per_gpu_train_batch_size: typing.Optional[int] = None per_gpu_eval_batch_size: typing.Optional[int] = None gradient_accumulation_steps: int = 1 eval_accumulation_steps: typing.Optional[int] = None eval_delay: typing.Optional[float] = 0 learning_rate: float = 5e-05 weight_decay: float = 0.0 adam_beta1: float = 0.9 adam_beta2: float = 0.999 adam_epsilon: float = 1e-08 max_grad_norm: float = 1.0 num_train_epochs: float = 3.0 max_steps: int = -1 lr_scheduler_type: typing.Union[transformers.trainer_utils.SchedulerType, str] = 'linear' warmup_ratio: float = 0.0 warmup_steps: int = 0 log_level: typing.Optional[str] = 'passive' log_level_replica: typing.Optional[str] = 'passive' log_on_each_node: bool = True logging_dir: typing.Optional[str] = None logging_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'steps' logging_first_step: bool = False logging_steps: int = 500 logging_nan_inf_filter: bool = True save_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'steps' save_steps: int = 500 save_total_limit: typing.Optional[int] = None save_on_each_node: bool = False no_cuda: bool = False seed: int = 42 data_seed: typing.Optional[int] = None jit_mode_eval: bool = False use_ipex: bool = False bf16: bool = False fp16: bool = False fp16_opt_level: str = 'O1' half_precision_backend: str = 'auto' bf16_full_eval: bool = False fp16_full_eval: bool = False tf32: typing.Optional[bool] = None local_rank: int = -1 xpu_backend: typing.Optional[str] = None tpu_num_cores: typing.Optional[int] = None tpu_metrics_debug: bool = False debug: str = '' dataloader_drop_last: bool = False eval_steps: typing.Optional[int] = None dataloader_num_workers: int = 0 past_index: int = -1 run_name: typing.Optional[str] = None disable_tqdm: typing.Optional[bool] = None remove_unused_columns: typing.Optional[bool] = True label_names: typing.Optional[typing.List[str]] = None load_best_model_at_end: typing.Optional[bool] = False metric_for_best_model: typing.Optional[str] = None greater_is_better: typing.Optional[bool] = None ignore_data_skip: bool = False sharded_ddp: str = '' fsdp: str = '' fsdp_min_num_params: int = 0 fsdp_transformer_layer_cls_to_wrap: typing.Optional[str] = None deepspeed: typing.Optional[str] = None label_smoothing_factor: float = 0.0 optim: typing.Union[transformers.training_args.OptimizerNames, str] = 'adamw_hf' adafactor: bool = False group_by_length: bool = False length_column_name: typing.Optional[str] = 'length' report_to: typing.Optional[typing.List[str]] = None ddp_find_unused_parameters: typing.Optional[bool] = None ddp_bucket_cap_mb: typing.Optional[int] = None dataloader_pin_memory: bool = True skip_memory_metrics: bool = True use_legacy_prediction_loop: bool = False push_to_hub: bool = False resume_from_checkpoint: typing.Optional[str] = None hub_model_id: typing.Optional[str] = None hub_strategy: typing.Union[transformers.trainer_utils.HubStrategy, str] = 'every_save' hub_token: typing.Optional[str] = None hub_private_repo: bool = False gradient_checkpointing: bool = False include_inputs_for_metrics: bool = False fp16_backend: str = 'auto' push_to_hub_model_id: typing.Optional[str] = None push_to_hub_organization: typing.Optional[str] = None push_to_hub_token: typing.Optional[str] = None mp_parameters: str = '' auto_find_batch_size: bool = False full_determinism: bool = False torchdynamo: typing.Optional[str] = None ray_scope: typing.Optional[str] = 'last' )

get_process_log_level

< >

( )

Returns the log level to be used depending on whether this process is the main process of node 0, main process of node non-0, or a non-main process.

For the main process the log level defaults to logging.INFO unless overridden by log_level argument.

For the replica processes the log level defaults to logging.WARNING unless overridden by log_level_replica argument.

The choice between the main and replica process settings is made according to the return value of should_log.

get_warmup_steps

< >

( num_training_steps: int )

Get number of steps used for a linear warmup.

main_process_first

< >

( local = True desc = 'work' )

Parameters

  • local (bool, optional, defaults to True) — if True first means process of rank 0 of each node if False first means process of rank 0 of node rank 0 In multi-node environment with a shared filesystem you most likely will want to use local=False so that only the main process of the first node will do the processing. If however, the filesystem is not shared, then the main process of each node will need to do the processing, which is the default behavior.
  • desc (str, optional, defaults to "work") — a work description to be used in debug logs

A context manager for torch distributed environment where on needs to do something on the main process, while blocking replicas, and when it’s finished releasing the replicas.

One such use is for datasets’s map feature which to be efficient should be run once on the main process, which upon completion saves a cached version of results and which then automatically gets loaded by the replicas.

to_dict

< >

( )

Serializes this instance while replace Enum by their values (for JSON serialization support). It obfuscates the token values by removing their value.

to_json_string

< >

( )

Serializes this instance to a JSON string.

to_sanitized_dict

< >

( )

Sanitized serialization to use with TensorBoard’s hparams

Seq2SeqTrainingArguments

class transformers.Seq2SeqTrainingArguments

< >

( output_dir: str overwrite_output_dir: bool = False do_train: bool = False do_eval: bool = False do_predict: bool = False evaluation_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'no' prediction_loss_only: bool = False per_device_train_batch_size: int = 8 per_device_eval_batch_size: int = 8 per_gpu_train_batch_size: typing.Optional[int] = None per_gpu_eval_batch_size: typing.Optional[int] = None gradient_accumulation_steps: int = 1 eval_accumulation_steps: typing.Optional[int] = None eval_delay: typing.Optional[float] = 0 learning_rate: float = 5e-05 weight_decay: float = 0.0 adam_beta1: float = 0.9 adam_beta2: float = 0.999 adam_epsilon: float = 1e-08 max_grad_norm: float = 1.0 num_train_epochs: float = 3.0 max_steps: int = -1 lr_scheduler_type: typing.Union[transformers.trainer_utils.SchedulerType, str] = 'linear' warmup_ratio: float = 0.0 warmup_steps: int = 0 log_level: typing.Optional[str] = 'passive' log_level_replica: typing.Optional[str] = 'passive' log_on_each_node: bool = True logging_dir: typing.Optional[str] = None logging_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'steps' logging_first_step: bool = False logging_steps: int = 500 logging_nan_inf_filter: bool = True save_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'steps' save_steps: int = 500 save_total_limit: typing.Optional[int] = None save_on_each_node: bool = False no_cuda: bool = False seed: int = 42 data_seed: typing.Optional[int] = None jit_mode_eval: bool = False use_ipex: bool = False bf16: bool = False fp16: bool = False fp16_opt_level: str = 'O1' half_precision_backend: str = 'auto' bf16_full_eval: bool = False fp16_full_eval: bool = False tf32: typing.Optional[bool] = None local_rank: int = -1 xpu_backend: typing.Optional[str] = None tpu_num_cores: typing.Optional[int] = None tpu_metrics_debug: bool = False debug: str = '' dataloader_drop_last: bool = False eval_steps: typing.Optional[int] = None dataloader_num_workers: int = 0 past_index: int = -1 run_name: typing.Optional[str] = None disable_tqdm: typing.Optional[bool] = None remove_unused_columns: typing.Optional[bool] = True label_names: typing.Optional[typing.List[str]] = None load_best_model_at_end: typing.Optional[bool] = False metric_for_best_model: typing.Optional[str] = None greater_is_better: typing.Optional[bool] = None ignore_data_skip: bool = False sharded_ddp: str = '' fsdp: str = '' fsdp_min_num_params: int = 0 fsdp_transformer_layer_cls_to_wrap: typing.Optional[str] = None deepspeed: typing.Optional[str] = None label_smoothing_factor: float = 0.0 optim: typing.Union[transformers.training_args.OptimizerNames, str] = 'adamw_hf' adafactor: bool = False group_by_length: bool = False length_column_name: typing.Optional[str] = 'length' report_to: typing.Optional[typing.List[str]] = None ddp_find_unused_parameters: typing.Optional[bool] = None ddp_bucket_cap_mb: typing.Optional[int] = None dataloader_pin_memory: bool = True skip_memory_metrics: bool = True use_legacy_prediction_loop: bool = False push_to_hub: bool = False resume_from_checkpoint: typing.Optional[str] = None hub_model_id: typing.Optional[str] = None hub_strategy: typing.Union[transformers.trainer_utils.HubStrategy, str] = 'every_save' hub_token: typing.Optional[str] = None hub_private_repo: bool = False gradient_checkpointing: bool = False include_inputs_for_metrics: bool = False fp16_backend: str = 'auto' push_to_hub_model_id: typing.Optional[str] = None push_to_hub_organization: typing.Optional[str] = None push_to_hub_token: typing.Optional[str] = None mp_parameters: str = '' auto_find_batch_size: bool = False full_determinism: bool = False torchdynamo: typing.Optional[str] = None ray_scope: typing.Optional[str] = 'last' sortish_sampler: bool = False predict_with_generate: bool = False generation_max_length: typing.Optional[int] = None generation_num_beams: typing.Optional[int] = None )

Parameters

  • sortish_sampler (bool, optional, defaults to False) — Whether to use a sortish sampler or not. Only possible if the underlying datasets are Seq2SeqDataset for now but will become generally available in the near future.

    It sorts the inputs according to lengths in order to minimize the padding size, with a bit of randomness for the training set.

  • predict_with_generate (bool, optional, defaults to False) — Whether to use generate to calculate generative metrics (ROUGE, BLEU).
  • generation_max_length (int, optional) — The max_length to use on each evaluation loop when predict_with_generate=True. Will default to the max_length value of the model configuration.
  • generation_num_beams (int, optional) — The num_beams to use on each evaluation loop when predict_with_generate=True. Will default to the num_beams value of the model configuration.

TrainingArguments(output_dir: str, overwrite_output_dir: bool = False, do_train: bool = False, do_eval: bool = False, do_predict: bool = False, evaluation_strategy: Union[transformers.trainer_utils.IntervalStrategy, str] = ‘no’, prediction_loss_only: bool = False, per_device_train_batch_size: int = 8, per_device_eval_batch_size: int = 8, per_gpu_train_batch_size: Union[int, NoneType] = None, per_gpu_eval_batch_size: Union[int, NoneType] = None, gradient_accumulation_steps: int = 1, eval_accumulation_steps: Union[int, NoneType] = None, eval_delay: Union[float, NoneType] = 0, learning_rate: float = 5e-05, weight_decay: float = 0.0, adam_beta1: float = 0.9, adam_beta2: float = 0.999, adam_epsilon: float = 1e-08, max_grad_norm: float = 1.0, num_train_epochs: float = 3.0, max_steps: int = -1, lr_scheduler_type: Union[transformers.trainer_utils.SchedulerType, str] = ‘linear’, warmup_ratio: float = 0.0, warmup_steps: int = 0, log_level: Union[str, NoneType] = ‘passive’, log_level_replica: Union[str, NoneType] = ‘passive’, log_on_each_node: bool = True, logging_dir: Union[str, NoneType] = None, logging_strategy: Union[transformers.trainer_utils.IntervalStrategy, str] = ‘steps’, logging_first_step: bool = False, logging_steps: int = 500, logging_nan_inf_filter: bool = True, save_strategy: Union[transformers.trainer_utils.IntervalStrategy, str] = ‘steps’, save_steps: int = 500, save_total_limit: Union[int, NoneType] = None, save_on_each_node: bool = False, no_cuda: bool = False, seed: int = 42, data_seed: Union[int, NoneType] = None, jit_mode_eval: bool = False, use_ipex: bool = False, bf16: bool = False, fp16: bool = False, fp16_opt_level: str = ‘O1’, half_precision_backend: str = ‘auto’, bf16_full_eval: bool = False, fp16_full_eval: bool = False, tf32: Union[bool, NoneType] = None, local_rank: int = -1, xpu_backend: Union[str, NoneType] = None, tpu_num_cores: Union[int, NoneType] = None, tpu_metrics_debug: bool = False, debug: str = ”, dataloader_drop_last: bool = False, eval_steps: Union[int, NoneType] = None, dataloader_num_workers: int = 0, past_index: int = -1, run_name: Union[str, NoneType] = None, disable_tqdm: Union[bool, NoneType] = None, remove_unused_columns: Union[bool, NoneType] = True, label_names: Union[List[str], NoneType] = None, load_best_model_at_end: Union[bool, NoneType] = False, metric_for_best_model: Union[str, NoneType] = None, greater_is_better: Union[bool, NoneType] = None, ignore_data_skip: bool = False, sharded_ddp: str = ”, fsdp: str = ”, fsdp_min_num_params: int = 0, fsdp_transformer_layer_cls_to_wrap: Union[str, NoneType] = None, deepspeed: Union[str, NoneType] = None, label_smoothing_factor: float = 0.0, optim: Union[transformers.training_args.OptimizerNames, str] = ‘adamw_hf’, adafactor: bool = False, group_by_length: bool = False, length_column_name: Union[str, NoneType] = ‘length’, report_to: Union[List[str], NoneType] = None, ddp_find_unused_parameters: Union[bool, NoneType] = None, ddp_bucket_cap_mb: Union[int, NoneType] = None, dataloader_pin_memory: bool = True, skip_memory_metrics: bool = True, use_legacy_prediction_loop: bool = False, push_to_hub: bool = False, resume_from_checkpoint: Union[str, NoneType] = None, hub_model_id: Union[str, NoneType] = None, hub_strategy: Union[transformers.trainer_utils.HubStrategy, str] = ‘every_save’, hub_token: Union[str, NoneType] = None, hub_private_repo: bool = False, gradient_checkpointing: bool = False, include_inputs_for_metrics: bool = False, fp16_backend: str = ‘auto’, push_to_hub_model_id: Union[str, NoneType] = None, push_to_hub_organization: Union[str, NoneType] = None, push_to_hub_token: Union[str, NoneType] = None, mp_parameters: str = ”, auto_find_batch_size: bool = False, full_determinism: bool = False, torchdynamo: Union[str, NoneType] = None, ray_scope: Union[str, NoneType] = ‘last’)

Checkpoints

By default, Trainer will save all checkpoints in the output_dir you set in the TrainingArguments you are using. Those will go in subfolder named checkpoint-xxx with xxx being the step at which the training was at.

Resuming training from a checkpoint can be done when calling Trainer.train() with either:

  • resume_from_checkpoint=True which will resume training from the latest checkpoint
  • resume_from_checkpoint=checkpoint_dir which will resume training from the specific checkpoint in the directory passed.

In addition, you can easily save your checkpoints on the Model Hub when using push_to_hub=True. By default, all the models saved in intermediate checkpoints are saved in different commits, but not the optimizer state. You can adapt the hub-strategy value of your TrainingArguments to either:

  • "checkpoint": the latest checkpoint is also pushed in a subfolder named last-checkpoint, allowing you to resume training easily with trainer.train(resume_from_checkpoint="output_dir/last-checkpoint").
  • "all_checkpoints": all checkpoints are pushed like they appear in the output folder (so you will get one checkpoint folder per folder in your final repository)

Logging

By default Trainer will use logging.INFO for the main process and logging.WARNING for the replicas if any.

These defaults can be overridden to use any of the 5 logging levels with TrainingArguments’s arguments:

  • log_level - for the main process
  • log_level_replica - for the replicas

Further, if TrainingArguments’s log_on_each_node is set to False only the main node will use the log level settings for its main process, all other nodes will use the log level settings for replicas.

Note that Trainer is going to set transformers’s log level separately for each node in its Trainer.__init__(). So you may want to set this sooner (see the next example) if you tap into other transformers functionality before creating the Trainer object.

Here is an example of how this can be used in an application:

[...]
logger = logging.getLogger(__name__)

# Setup logging
logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%m/%d/%Y %H:%M:%S",
    handlers=[logging.StreamHandler(sys.stdout)],
)

# set the main code and the modules it uses to the same log-level according to the node
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)

trainer = Trainer(...)

And then if you only want to see warnings on the main node and all other nodes to not print any most likely duplicated warnings you could run it as:

my_app.py ... --log_level warning --log_level_replica error

In the multi-node environment if you also don’t want the logs to repeat for each node’s main process, you will want to change the above to:

my_app.py ... --log_level warning --log_level_replica error --log_on_each_node 0

and then only the main process of the first node will log at the “warning” level, and all other processes on the main node and all processes on other nodes will log at the “error” level.

If you need your application to be as quiet as possible you could do:

my_app.py ... --log_level error --log_level_replica error --log_on_each_node 0

(add --log_on_each_node 0 if on multi-node environment)

Randomness

When resuming from a checkpoint generated by Trainer all efforts are made to restore the python, numpy and pytorch RNG states to the same states as they were at the moment of saving that checkpoint, which should make the “stop and resume” style of training as close as possible to non-stop training.

However, due to various default non-deterministic pytorch settings this might not fully work. If you want full determinism please refer to Controlling sources of randomness. As explained in the document, that some of those settings that make things deterministic (.e.g., torch.backends.cudnn.deterministic) may slow things down, therefore this can’t be done by default, but you can enable those yourself if needed.

Specific GPUs Selection

Let’s discuss how you can tell your program which GPUs are to be used and in what order.

When using DistributedDataParallel to use only a subset of your GPUs, you simply specify the number of GPUs to use. For example, if you have 4 GPUs, but you wish to use the first 2 you can do:

python -m torch.distributed.launch --nproc_per_node=2  trainer-program.py ...

if you have either accelerate or deepspeed installed you can also accomplish the same by using one of:

accelerate launch --num_processes 2 trainer-program.py ...
deepspeed --num_gpus 2 trainer-program.py ...

You don’t need to use the Accelerate or the Deepspeed integration features to use these launchers.

Until now you were able to tell the program how many GPUs to use. Now let’s discuss how to select specific GPUs and control their order.

The following environment variables help you control which GPUs to use and their order.

CUDA_VISIBLE_DEVICES

If you have multiple GPUs and you’d like to use only 1 or a few of those GPUs, set the environment variable CUDA_VISIBLE_DEVICES to a list of the GPUs to be used.

For example, let’s say you have 4 GPUs: 0, 1, 2 and 3. To run only on the physical GPUs 0 and 2, you can do:

CUDA_VISIBLE_DEVICES=0,2 python -m torch.distributed.launch trainer-program.py ...

So now pytorch will see only 2 GPUs, where your physical GPUs 0 and 2 are mapped to cuda:0 and cuda:1 correspondingly.

You can even change their order:

CUDA_VISIBLE_DEVICES=2,0 python -m torch.distributed.launch trainer-program.py ...

Here your physical GPUs 0 and 2 are mapped to cuda:1 and cuda:0 correspondingly.

The above examples were all for DistributedDataParallel use pattern, but the same method works for DataParallel as well:

CUDA_VISIBLE_DEVICES=2,0 python trainer-program.py ...

To emulate an environment without GPUs simply set this environment variable to an empty value like so:

CUDA_VISIBLE_DEVICES= python trainer-program.py ...

As with any environment variable you can, of course, export those instead of adding these to the command line, as in:

export CUDA_VISIBLE_DEVICES=0,2
python -m torch.distributed.launch trainer-program.py ...

but this approach can be confusing since you may forget you set up the environment variable earlier and not understand why the wrong GPUs are used. Therefore, it’s a common practice to set the environment variable just for a specific run on the same command line as it’s shown in most examples of this section.

CUDA_DEVICE_ORDER

There is an additional environment variable CUDA_DEVICE_ORDER that controls how the physical devices are ordered. The two choices are:

  1. ordered by PCIe bus IDs (matches nvidia-smi’s order) - this is the default.
export CUDA_DEVICE_ORDER=PCI_BUS_ID
  1. ordered by GPU compute capabilities
export CUDA_DEVICE_ORDER=FASTEST_FIRST

Most of the time you don’t need to care about this environment variable, but it’s very helpful if you have a lopsided setup where you have an old and a new GPUs physically inserted in such a way so that the slow older card appears to be first. One way to fix that is to swap the cards. But if you can’t swap the cards (e.g., if the cooling of the devices gets impacted) then setting CUDA_DEVICE_ORDER=FASTEST_FIRST will always put the newer faster card first. It’ll be somewhat confusing though since nvidia-smi will still report them in the PCIe order.

The other solution to swapping the order is to use:

export CUDA_VISIBLE_DEVICES=1,0

In this example we are working with just 2 GPUs, but of course the same would apply to as many GPUs as your computer has.

Also if you do set this environment variable it’s the best to set it in your ~/.bashrc file or some other startup config file and forget about it.

Trainer Integrations

The Trainer has been extended to support libraries that may dramatically improve your training time and fit much bigger models.

Currently it supports third party solutions, DeepSpeed, PyTorch FSDP and FairScale, which implement parts of the paper ZeRO: Memory Optimizations Toward Training Trillion Parameter Models, by Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, Yuxiong He.

This provided support is new and experimental as of this writing. While the support for DeepSpeed and PyTorch FSDP is active and we welcome issues around it, we don’t support the FairScale integration anymore since it has been integrated in PyTorch main (see the PyTorch FSDP integration)

CUDA Extension Installation Notes

As of this writing, both FairScale and Deepspeed require compilation of CUDA C++ code, before they can be used.

While all installation issues should be dealt with through the corresponding GitHub Issues of FairScale and Deepspeed, there are a few common issues that one may encounter while building any PyTorch extension that needs to build CUDA extensions.

Therefore, if you encounter a CUDA-related build issue while doing one of the following or both:

pip install fairscale
pip install deepspeed

please, read the following notes first.

In these notes we give examples for what to do when pytorch has been built with CUDA 10.2. If your situation is different remember to adjust the version number to the one you are after.

Possible problem #1

While, Pytorch comes with its own CUDA toolkit, to build these two projects you must have an identical version of CUDA installed system-wide.

For example, if you installed pytorch with cudatoolkit==10.2 in the Python environment, you also need to have CUDA 10.2 installed system-wide.

The exact location may vary from system to system, but /usr/local/cuda-10.2 is the most common location on many Unix systems. When CUDA is correctly set up and added to the PATH environment variable, one can find the installation location by doing:

which nvcc

If you don’t have CUDA installed system-wide, install it first. You will find the instructions by using your favorite search engine. For example, if you’re on Ubuntu you may want to search for: ubuntu cuda 10.2 install.

Possible problem #2

Another possible common problem is that you may have more than one CUDA toolkit installed system-wide. For example you may have:

/usr/local/cuda-10.2
/usr/local/cuda-11.0

Now, in this situation you need to make sure that your PATH and LD_LIBRARY_PATH environment variables contain the correct paths to the desired CUDA version. Typically, package installers will set these to contain whatever the last version was installed. If you encounter the problem, where the package build fails because it can’t find the right CUDA version despite you having it installed system-wide, it means that you need to adjust the 2 aforementioned environment variables.

First, you may look at their contents:

echo $PATH
echo $LD_LIBRARY_PATH

so you get an idea of what is inside.

It’s possible that LD_LIBRARY_PATH is empty.

PATH lists the locations of where executables can be found and LD_LIBRARY_PATH is for where shared libraries are to looked for. In both cases, earlier entries have priority over the later ones. : is used to separate multiple entries.

Now, to tell the build program where to find the specific CUDA toolkit, insert the desired paths to be listed first by doing:

export PATH=/usr/local/cuda-10.2/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-10.2/lib64:$LD_LIBRARY_PATH

Note that we aren’t overwriting the existing values, but prepending instead.

Of course, adjust the version number, the full path if need be. Check that the directories you assign actually do exist. lib64 sub-directory is where the various CUDA .so objects, like libcudart.so reside, it’s unlikely that your system will have it named differently, but if it is adjust it to reflect your reality.

Possible problem #3

Some older CUDA versions may refuse to build with newer compilers. For example, you my have gcc-9 but it wants gcc-7.

There are various ways to go about it.

If you can install the latest CUDA toolkit it typically should support the newer compiler.

Alternatively, you could install the lower version of the compiler in addition to the one you already have, or you may already have it but it’s not the default one, so the build system can’t see it. If you have gcc-7 installed but the build system complains it can’t find it, the following might do the trick:

sudo ln -s /usr/bin/gcc-7  /usr/local/cuda-10.2/bin/gcc
sudo ln -s /usr/bin/g++-7  /usr/local/cuda-10.2/bin/g++

Here, we are making a symlink to gcc-7 from /usr/local/cuda-10.2/bin/gcc and since /usr/local/cuda-10.2/bin/ should be in the PATH environment variable (see the previous problem’s solution), it should find gcc-7 (and g++7) and then the build will succeed.

As always make sure to edit the paths in the example to match your situation.

FairScale

This integration is not supported anymore, we recommend you either use DeepSpeed or PyTorch FSDP.

By integrating FairScale the Trainer provides support for the following features from the ZeRO paper:

  1. Optimizer State Sharding
  2. Gradient Sharding
  3. Model Parameters Sharding (new and very experimental)
  4. CPU offload (new and very experimental)

You will need at least two GPUs to use this feature.

Installation:

Install the library via pypi:

pip install fairscale

or via transformersextras:

pip install transformers[fairscale]

(available starting from transformers==4.6.0) or find more details on the FairScale’s GitHub page.

If you’re still struggling with the build, first make sure to read CUDA Extension Installation Notes.

If it’s still not resolved the build issue, here are a few more ideas.

fairscale seems to have an issue with the recently introduced by pip build isolation feature. If you have a problem with it, you may want to try one of:

pip install fairscale --no-build-isolation .

or:

git clone https://github.com/facebookresearch/fairscale/
cd fairscale
rm -r dist build
python setup.py bdist_wheel
pip uninstall -y fairscale
pip install dist/fairscale-*.whl

fairscale also has issues with building against pytorch-nightly, so if you use it you may have to try one of:

pip uninstall -y fairscale; pip install fairscale --pre \
-f https://download.pytorch.org/whl/nightly/cu110/torch_nightly \
--no-cache --no-build-isolation

or:

pip install -v --disable-pip-version-check . \
-f https://download.pytorch.org/whl/nightly/cu110/torch_nightly --pre

Of course, adjust the urls to match the cuda version you use.

If after trying everything suggested you still encounter build issues, please, proceed with the GitHub Issue of FairScale.

Usage:

To use the first version of Sharded data-parallelism, add --sharded_ddp simple to the command line arguments, and make sure you have added the distributed launcher -m torch.distributed.launch --nproc_per_node=NUMBER_OF_GPUS_YOU_HAVE if you haven’t been using it already.

For example here is how you could use it for run_translation.py with 2 GPUs:

python -m torch.distributed.launch --nproc_per_node=2 examples/pytorch/translation/run_translation.py \
--model_name_or_path t5-small --per_device_train_batch_size 1   \
--output_dir output_dir --overwrite_output_dir \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro \
--fp16 --sharded_ddp simple

Notes:

  • This feature requires distributed training (so multiple GPUs).
  • It is not implemented for TPUs.
  • It works with --fp16 too, to make things even faster.
  • One of the main benefits of enabling --sharded_ddp simple is that it uses a lot less GPU memory, so you should be able to use significantly larger batch sizes using the same hardware (e.g. 3x and even bigger) which should lead to significantly shorter training time.
  1. To use the second version of Sharded data-parallelism, add --sharded_ddp zero_dp_2 or --sharded_ddp zero_dp_3 to the command line arguments, and make sure you have added the distributed launcher -m torch.distributed.launch --nproc_per_node=NUMBER_OF_GPUS_YOU_HAVE if you haven’t been using it already.

For example here is how you could use it for run_translation.py with 2 GPUs:

python -m torch.distributed.launch --nproc_per_node=2 examples/pytorch/translation/run_translation.py \
--model_name_or_path t5-small --per_device_train_batch_size 1   \
--output_dir output_dir --overwrite_output_dir \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro \
--fp16 --sharded_ddp zero_dp_2

zero_dp_2 is an optimized version of the simple wrapper, while zero_dp_3 fully shards model weights, gradients and optimizer states.

Both are compatible with adding cpu_offload to enable ZeRO-offload (activate it like this: --sharded_ddp "zero_dp_2 cpu_offload").

Notes:

  • This feature requires distributed training (so multiple GPUs).
  • It is not implemented for TPUs.
  • It works with --fp16 too, to make things even faster.
  • The cpu_offload additional option requires --fp16.
  • This is an area of active development, so make sure you have a source install of fairscale to use this feature as some bugs you encounter may have been fixed there already.

Known caveats:

  • This feature is incompatible with --predict_with_generate in the run_translation.py script.
  • Using --sharded_ddp zero_dp_3 requires wrapping each layer of the model in the special container FullyShardedDataParallelism of fairscale. It should be used with the option auto_wrap if you are not doing this yourself: --sharded_ddp "zero_dp_3 auto_wrap".

PyTorch Fully Sharded Data parallel

To accelerate training huge models on larger batch sizes, we can use a fully sharded data parallel model. This type of data parallel paradigm enables fitting more data and larger models by sharding the optimizer states, gradients and parameters. To read more about it and the benefits, check out the Fully Sharded Data Parallel blog. We have integrated the latest PyTorch’s Fully Sharded Data Parallel (FSDP) training feature. All you need to do is enable it through the config.

Required PyTorch version for FSDP support: PyTorch Nightly (or 1.12.0 if you read this after it has been released) as the model saving with FSDP activated is only available with recent fixes.

Usage:

  • Make sure you have added the distributed launcher -m torch.distributed.launch --nproc_per_node=NUMBER_OF_GPUS_YOU_HAVE if you haven’t been using it already.

  • Sharding Strategy:

    • FULL_SHARD : Shards optimizer states + gradients + model parameters across data parallel workers/GPUs. For this, add --fsdp full_shard to the command line arguments.
    • SHARD_GRAD_OP : Shards optimizer states + gradients across data parallel workers/GPUs. For this, add --fsdp shard_grad_op to the command line arguments.
    • NO_SHARD : No sharding. For this, add --fsdp no_shard to the command line arguments.
  • To offload the parameters and gradients to the CPU, add --fsdp "full_shard offload" or --fsdp "shard_grad_op offload" to the command line arguments.

  • To automatically recursively wrap layers with FSDP using default_auto_wrap_policy, add --fsdp "full_shard auto_wrap" or --fsdp "shard_grad_op auto_wrap" to the command line arguments.

  • To enable both CPU offloading and auto wrapping, add --fsdp "full_shard offload auto_wrap" or --fsdp "shard_grad_op offload auto_wrap" to the command line arguments.

  • If auto wrapping is enabled, you can either use transformer based auto wrap policy or size based auto wrap policy.

    • For transformer based auto wrap policy, please add --fsdp_transformer_layer_cls_to_wrap <value> to command line arguments. This specifies the transformer layer class name (case-sensitive) to wrap ,e.g, BertLayer, GPTJBlock, T5Block … This is important because submodules that share weights (e.g., embedding layer) should not end up in different FSDP wrapped units. Using this policy, wrapping happens for each block containing Multi-Head Attention followed by couple of MLP layers. Remaining layers including the shared embeddings are conviniently wrapped in same outermost FSDP unit. Therefore, use this for transformer based models.
    • For size based auto wrap policy, please add --fsdp_min_num_params <number> to command line arguments. It specifies FSDP’s minimum number of parameters for auto wrapping.

Few caveats to be aware of

  • Mixed precision is currently not supported with FSDP as we wait for PyTorch to fix support for it. More details in this issues.
  • FSDP currently doesn’t support multiple parameter groups. More details mentioned in this issue (The original model parameters' .grads are not set, meaning that they cannot be optimized separately (which is why we cannot support multiple parameter groups)).

Sections that were moved:

[ DeepSpeed | Installation | Deployment with multiple GPUs | Deployment with one GPU | Deployment in Notebooks | Configuration | Passing Configuration | Shared Configuration | ZeRO | ZeRO-2 Config | ZeRO-3 Config | NVMe Support | ZeRO-2 vs ZeRO-3 Performance | ZeRO-2 Example | ZeRO-3 Example | Optimizer | Scheduler | fp32 Precision | Automatic Mixed Precision | Batch Size | Gradient Accumulation | Gradient Clipping | Getting The Model Weights Out ]