Hugging Face's logo

Accelerate documentation

Accelerator

You are viewing v0.23.0 version. A newer version v1.1.0 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Accelerator

The Accelerator is the main class provided by 🤗 Accelerate. It serves at the main entry point for the API.

Quick adaptation of your code

To quickly adapt your script to work on any kind of setup with 🤗 Accelerate just:

  1. Initialize an Accelerator object (that we will call accelerator throughout this page) as early as possible in your script.
  2. Pass your dataloader(s), model(s), optimizer(s), and scheduler(s) to the prepare() method.
  3. Remove all the .cuda() or .to(device) from your code and let the accelerator handle the device placement for you.

Step three is optional, but considered a best practice.

  1. Replace loss.backward() in your code with accelerator.backward(loss)
  2. Gather your predictions and labels before storing them or using them for metric computation using gather()

Step five is mandatory when using distributed evaluation

In most cases this is all that is needed. The next section lists a few more advanced use cases and nice features you should search for and replace by the corresponding methods of your accelerator:

Advanced recommendations

Printing

print statements should be replaced by print() to be printed once per process:

- print("My thing I want to print!")
+ accelerator.print("My thing I want to print!")

Executing processes

Once on a single server

For statements that should be executed once per server, use is_local_main_process:

if accelerator.is_local_main_process:
    do_thing_once_per_server()

A function can be wrapped using the on_local_main_process() function to achieve the same behavior on a function’s execution:

@accelerator.on_local_main_process
def do_my_thing():
    "Something done once per server"
    do_thing_once_per_server()

Only ever once across all servers

For statements that should only ever be executed once, use is_main_process:

if accelerator.is_main_process:
    do_thing_once()

A function can be wrapped using the on_main_process() function to achieve the same behavior on a function’s execution:

@accelerator.on_main_process
def do_my_thing():
    "Something done once per server"
    do_thing_once()

On specific processes

If a function should be ran on a specific overall or local process index, there are similar decorators to achieve this:

@accelerator.on_local_process(local_process_idx=0)
def do_my_thing():
    "Something done on process index 0 on each server"
    do_thing_on_index_zero_on_each_server()
@accelerator.on_process(process_index=0)
def do_my_thing():
    "Something done on process index 0"
    do_thing_on_index_zero()

Synchronicity control

Use wait_for_everyone() to make sure all processes join that point before continuing. (Useful before a model save for instance).

Saving and loading 

model = MyModel()
model = accelerator.prepare(model)

Use save_model() instead of torch.save to save a model. It will remove all model wrappers added during the distributed process, get the state_dict of the model and save it. The state_dict will be in the same precision as the model being trained.

- torch.save(state_dict, "my_state.pkl")
+ accelerator.save_model(model, save_directory)

save_model() can also save a model into sharded checkpoints or with safetensors format. Here is an example: 

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

🤗 Transformers models

If you are using models from the 🤗 Transformers library, you can use the .save_pretrained() method.

from transformers import AutoModel

model = AutoModel.from_pretrained("bert-base-cased")
model = accelerator.prepare(model)

# ...fine-tune with PyTorch...

unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
    "path/to/my_model_directory",
    is_main_process=accelerator.is_main_process,
    save_function=accelerator.save,
)

This will ensure your model stays compatible with other 🤗 Transformers functionality like the .from_pretrained() method.

from transformers import AutoModel

model = AutoModel.from_pretrained("path/to/my_model_directory")

Operations

Use clipgrad_norm() instead of torch.nn.utils.clip_grad_norm_ and clipgrad_value() instead of torch.nn.utils.clip_grad_value

Gradient Accumulation

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

- accelerator = Accelerator()
+ accelerator = Accelerator(gradient_accumulation_steps=2)

  for (input, label) in training_dataloader:
+     with accelerator.accumulate(model):
          predictions = model(input)
          loss = loss_function(predictions, labels)
          accelerator.backward(loss)
          optimizer.step()
          scheduler.step()
          optimizer.zero_grad()

GradientAccumulationPlugin

class accelerate.utils.GradientAccumulationPlugin

< >

( num_steps: int = None adjust_scheduler: bool = True sync_with_dataloader: bool = True )

A plugin to configure gradient accumulation behavior.

Instead of passing gradient_accumulation_steps you can instantiate a GradientAccumulationPlugin and pass it to the Accelerator’s __init__ as gradient_accumulation_plugin. You can only pass either one of gradient_accumulation_plugin or gradient_accumulation_steps passing both will raise an error.

from accelerate.utils import GradientAccumulationPlugin

gradient_accumulation_plugin = GradientAccumulationPlugin(num_steps=2)
- accelerator = Accelerator()
+ accelerator = Accelerator(gradient_accumulation_plugin=gradient_accumulation_plugin)

In addition to the number of steps, this also lets you configure whether or not you adjust your learning rate scheduler to account for the change in steps due to accumulation.

Overall API documentation:

class accelerate.Accelerator

< >

( device_placement: bool = True split_batches: bool = False mixed_precision: PrecisionType | str | None = None gradient_accumulation_steps: int = 1 cpu: bool = False deepspeed_plugin: DeepSpeedPlugin | None = None fsdp_plugin: FullyShardedDataParallelPlugin | None = None megatron_lm_plugin: MegatronLMPlugin | None = None rng_types: list[str | RNGType] | None = None log_with: str | LoggerType | GeneralTracker | list[str | LoggerType | GeneralTracker] | None = None project_dir: str | os.PathLike | None = None project_config: ProjectConfiguration | None = None gradient_accumulation_plugin: GradientAccumulationPlugin | None = None dispatch_batches: bool | None = None even_batches: bool = True step_scheduler_with_optimizer: bool = True kwargs_handlers: list[KwargsHandler] | None = None dynamo_backend: DynamoBackend | str | None = None )

Parameters

  • device_placement (bool, optional, defaults to True) — Whether or not the accelerator should put objects on device (tensors yielded by the dataloader, model, etc…).
  • split_batches (bool, optional, defaults to False) — Whether or not the accelerator should split the batches yielded by the dataloaders across the devices. If True the actual batch size used will be the same on any kind of distributed processes, but it must be a round multiple of the num_processes you are using. If False, actual batch size used will be the one set in your script multiplied by the number of processes.
  • mixed_precision (str, optional) — Whether or not to use mixed precision training. Choose from ‘no’,‘fp16’,‘bf16 or ‘fp8’. Will default to the value in the environment variable ACCELERATE_MIXED_PRECISION, which will use the default value in the accelerate config of the current system or the flag passed with the accelerate.launch command. ‘fp8’ requires the installation of transformers-engine.
  • gradient_accumulation_steps (int, optional, default to 1) — The number of steps that should pass before gradients are accumulated. A number > 1 should be combined with Accelerator.accumulate. If not passed, will default to the value in the environment variable ACCELERATE_GRADIENT_ACCUMULATION_STEPS. Can also be configured through a GradientAccumulationPlugin.
  • cpu (bool, optional) — Whether or not to force the script to execute on CPU. Will ignore GPU available if set to True and force the execution on one process only.
  • deepspeed_plugin (DeepSpeedPlugin, optional) — Tweak your DeepSpeed related args using this argument. This argument is optional and can be configured directly using accelerate config
  • fsdp_plugin (FullyShardedDataParallelPlugin, optional) — Tweak your FSDP related args using this argument. This argument is optional and can be configured directly using accelerate config
  • megatron_lm_plugin (MegatronLMPlugin, optional) — Tweak your MegatronLM related args using this argument. This argument is optional and can be configured directly using accelerate config
  • rng_types (list of str or RNGType) — The list of random number generators to synchronize at the beginning of each iteration in your prepared dataloaders. Should be one or several of:

    • "torch": the base torch random number generator
    • "cuda": the CUDA random number generator (GPU only)
    • "xla": the XLA random number generator (TPU only)
    • "generator": the torch.Generator of the sampler (or batch sampler if there is no sampler in your dataloader) or of the iterable dataset (if it exists) if the underlying dataset is of that type.

    Will default to ["torch"] for PyTorch versions <=1.5.1 and ["generator"] for PyTorch versions >= 1.6.

  • log_with (list of str, LoggerType or GeneralTracker, optional) — A list of loggers to be setup for experiment tracking. Should be one or several of:

    • "all"
    • "tensorboard"
    • "wandb"
    • "comet_ml" If "all" is selected, will pick up all available trackers in the environment and initialize them. Can also accept implementations of GeneralTracker for custom trackers, and can be combined with "all".
  • project_config (ProjectConfiguration, optional) — A configuration for how saving the state can be handled.
  • project_dir (str, os.PathLike, optional) — A path to a directory for storing data such as logs of locally-compatible loggers and potentially saved checkpoints.
  • dispatch_batches (bool, optional) — If set to True, the dataloader prepared by the Accelerator is only iterated through on the main process and then the batches are split and broadcast to each process. Will default to True for DataLoader whose underlying dataset is an IterableDataset, False otherwise.
  • even_batches (bool, optional, defaults to True) — If set to True, in cases where the total batch size across all processes does not exactly divide the dataset, samples at the start of the dataset will be duplicated so the batch can be divided equally among all workers.
  • step_scheduler_with_optimizer (bool, *optional, defaults to True) -- Set Trueif the learning rate scheduler is stepped at the same time as the optimizer,False` if only done under certain circumstances (at the end of each epoch, for instance).
  • kwargs_handlers (list[KwargHandler], optional) — A list of KwargHandler to customize how the objects related to distributed training or mixed precision are created. See kwargs for more information.
  • dynamo_backend (str or DynamoBackend, optional, defaults to "no") — Set to one of the possible dynamo backends to optimize your training with torch dynamo.
  • gradient_accumulation_plugin (GradientAccumulationPlugin, optional) — A configuration for how gradient accumulation should be handled, if more tweaking than just the gradient_accumulation_steps is needed.

Creates an instance of an accelerator for distributed training (on multi-GPU, TPU) or mixed precision training.

Available attributes:

  • device (torch.device) — The device to use.
  • distributed_type (DistributedType) — The distributed training configuration.
  • local_process_index (int) — The process index on the current machine.
  • mixed_precision (str) — The configured mixed precision mode.
  • num_processes (int) — The total number of processes used for training.
  • optimizer_step_was_skipped (bool) — Whether or not the optimizer update was skipped (because of gradient overflow in mixed precision), in which case the learning rate should not be changed.
  • process_index (int) — The overall index of the current process among all processes.
  • state (AcceleratorState) — The distributed setup state.
  • sync_gradients (bool) — Whether the gradients are currently being synced across all processes.
  • use_distributed (bool) — Whether the current configuration is for distributed training.

accumulate

< >

( *models )

Parameters

  • *models (list of torch.nn.Module) — PyTorch Modules that was prepared with Accelerator.prepare. Models passed to accumulate() will skip gradient syncing during backward pass in distributed training

A context manager that will lightly wrap around and perform gradient accumulation automatically

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator(gradient_accumulation_steps=1)
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)

>>> for input, output in dataloader:
...     with accelerator.accumulate(model):
...         outputs = model(input)
...         loss = loss_func(outputs)
...         loss.backward()
...         optimizer.step()
...         scheduler.step()
...         optimizer.zero_grad()

autocast

< >

( cache_enabled: bool = False autocast_handler: AutocastKwargs = None )

Will apply automatic mixed-precision inside the block inside this context manager, if it is enabled. Nothing different will happen otherwise.

A different autocast_handler can be passed in to override the one set in the Accelerator object. This is useful in blocks under autocast where you want to revert to fp32.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator(mixed_precision="fp16")
>>> with accelerator.autocast():
...     train()

backward

< >

( loss **kwargs )

Scales the gradients in accordance to the GradientAccumulationPlugin and calls the correct backward() based on the configuration.

Should be used in lieu of loss.backward().

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator(gradient_accumulation_steps=2)
>>> outputs = model(inputs)
>>> loss = loss_fn(outputs, labels)
>>> accelerator.backward(loss)

check_trigger

< >

( )

Checks if the internal trigger tensor has been set to 1 in any of the processes. If so, will return True and reset the trigger tensor to 0.

Note: Does not require wait_for_everyone()

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> # Assume later in the training script
>>> # `should_do_breakpoint` is a custom function to monitor when to break,
>>> # e.g. when the loss is NaN
>>> if should_do_breakpoint(loss):
...     accelerator.set_trigger()
>>> # Assume later in the training script
>>> if accelerator.check_trigger():
...     break

clear

< >

( )

Alias for Accelerate.free_memory, releases all references to the internal objects stored and call the garbage collector. You should call this method between two trainings with different models/optimizers.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> model, optimizer, scheduler = ...
>>> model, optimizer, scheduler = accelerator.prepare(model, optimizer, scheduler)
>>> accelerator.free_memory()
>>> del model, optimizer, scheduler

clip_grad_norm_

< >

( parameters max_norm norm_type = 2 ) → torch.Tensor

Returns

torch.Tensor

Total norm of the parameter gradients (viewed as a single vector).

Should be used in place of torch.nn.utils.clip_grad_norm_.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator(gradient_accumulation_steps=2)
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)

>>> for input, target in dataloader:
...     optimizer.zero_grad()
...     output = model(input)
...     loss = loss_func(output, target)
...     accelerator.backward(loss)
...     if accelerator.sync_gradients:
...         accelerator.clip_grad_norm_(model.parameters(), max_grad_norm)
...     optimizer.step()

clip_grad_value_

< >

( parameters clip_value )

Should be used in place of torch.nn.utils.clip_grad_value_.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator(gradient_accumulation_steps=2)
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)

>>> for input, target in dataloader:
...     optimizer.zero_grad()
...     output = model(input)
...     loss = loss_func(output, target)
...     accelerator.backward(loss)
...     if accelerator.sync_gradients:
...         accelerator.clip_grad_value_(model.parameters(), clip_value)
...     optimizer.step()

free_memory

< >

( )

Will release all references to the internal objects stored and call the garbage collector. You should call this method between two trainings with different models/optimizers. Also will reset Accelerator.step to 0.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> model, optimizer, scheduler = ...
>>> model, optimizer, scheduler = accelerator.prepare(model, optimizer, scheduler)
>>> accelerator.free_memory()
>>> del model, optimizer, scheduler

gather

< >

( tensor ) → torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor

Parameters

  • tensor (torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor) — The tensors to gather across all processes.

Returns

torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor

The gathered tensor(s). Note that the first dimension of the result is num_processes multiplied by the first dimension of the input tensors.

Gather the values in tensor across all processes and concatenate them on the first dimension. Useful to regroup the predictions from all processes when doing evaluation.

Note: This gather happens in all processes.

Example:

>>> # Assuming four processes
>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> process_tensor = torch.tensor([accelerator.process_index])
>>> gathered_tensor = accelerator.gather(process_tensor)
>>> gathered_tensor
tensor([0, 1, 2, 3])

gather_for_metrics

< >

( input_data )

Parameters

  • input (torch.Tensor, object, a nested tuple/list/dictionary of torch.Tensor, or a nested tuple/list/dictionary of object) — The tensors or objects for calculating metrics across all processes

Gathers input_data and potentially drops duplicates in the last batch if on a distributed system. Should be used for gathering the inputs and targets for metric calculation.

Example:

>>> # Assuming two processes, with a batch size of 5 on a dataset with 9 samples
>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> dataloader = torch.utils.data.DataLoader(range(9), batch_size=5)
>>> dataloader = accelerator.prepare(dataloader)
>>> batch = next(iter(dataloader))
>>> gathered_items = accelerator.gather_for_metrics(batch)
>>> len(gathered_items)
9

get_state_dict

< >

( model unwrap = True ) → dict

Parameters

  • model (torch.nn.Module) — A PyTorch model sent through Accelerator.prepare()
  • unwrap (bool, optional, defaults to True) — Whether to return the original underlying state_dict of model or to return the wrapped state_dict

Returns

dict

The state dictionary of the model potentially without full precision.

Returns the state dictionary of a model sent through Accelerator.prepare() potentially without full precision.

Example:

>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> net = torch.nn.Linear(2, 2)
>>> net = accelerator.prepare(net)
>>> state_dict = accelerator.get_state_dict(net)

get_tracker

< >

( name: str unwrap: bool = False ) → GeneralTracker

Parameters

  • name (str) — The name of a tracker, corresponding to the .name property.
  • unwrap (bool) — Whether to return the internal tracking mechanism or to return the wrapped tracker instead (recommended).

Returns

GeneralTracker

The tracker corresponding to name if it exists.

Returns a tracker from self.trackers based on name on the main process only.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator(log_with="tensorboard")
>>> accelerator.init_trackers("my_project")
>>> tensorboard_tracker = accelerator.get_tracker("tensorboard")

join_uneven_inputs

< >

( joinables even_batches = None )

Parameters

  • joinables (list[torch.distributed.algorithms.Joinable]) — A list of models or optimizers that subclass torch.distributed.algorithms.Joinable. Most commonly, a PyTorch Module that was prepared with Accelerator.prepare for DistributedDataParallel training.
  • even_batches (bool, optional) — If set, this will override the value of even_batches set in the Accelerator. If it is not provided, the default Accelerator value wil be used.

A context manager that facilitates distributed training or evaluation on uneven inputs, which acts as a wrapper around torch.distributed.algorithms.join. This is useful when the total batch size does not evenly divide the length of the dataset.

join_uneven_inputs is only supported for Distributed Data Parallel training on multiple GPUs. For any other configuration, this method will have no effect.

Overidding even_batches will not affect iterable-style data loaders.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator(even_batches=True)
>>> ddp_model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)

>>> with accelerator.join_uneven_inputs([ddp_model], even_batches=False):
...     for input, output in dataloader:
...         outputs = model(input)
...         loss = loss_func(outputs)
...         loss.backward()
...         optimizer.step()
...         optimizer.zero_grad()

load_state

< >

( input_dir: str = None **load_model_func_kwargs )

Parameters

  • input_dir (str or os.PathLike) — The name of the folder all relevant weights and states were saved in. Can be None if automatic_checkpoint_naming is used, and will pick up from the latest checkpoint.
  • load_model_func_kwargs (dict, optional) — Additional keyword arguments for loading model which can be passed to the underlying load function, such as optional arguments for DeepSpeed’s load_checkpoint function or a map_location to load the model and optimizer on.

Loads the current states of the model, optimizer, scaler, RNG generators, and registered objects.

Should only be used in conjunction with Accelerator.save_state(). If a file is not registered for checkpointing, it will not be loaded if stored in the directory.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> model, optimizer, lr_scheduler = ...
>>> model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
>>> accelerator.load_state("my_checkpoint")

local_main_process_first

< >

( )

Lets the local main process go inside a with block.

The other processes will enter the with block after the main process exits.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> with accelerator.local_main_process_first():
...     # This will be printed first by local process 0 then in a seemingly
...     # random order by the other processes.
...     print(f"This will be printed by process {accelerator.local_process_index}")

main_process_first

< >

( )

Lets the main process go first inside a with block.

The other processes will enter the with block after the main process exits.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> with accelerator.main_process_first():
...     # This will be printed first by process 0 then in a seemingly
...     # random order by the other processes.
...     print(f"This will be printed by process {accelerator.process_index}")

no_sync

< >

( model )

Parameters

  • model (torch.nn.Module) — PyTorch Module that was prepared with Accelerator.prepare

A context manager to disable gradient synchronizations across DDP processes by calling torch.nn.parallel.DistributedDataParallel.no_sync.

If model is not in DDP, this context manager does nothing

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> dataloader, model, optimizer = accelerator.prepare(dataloader, model, optimizer)
>>> input_a = next(iter(dataloader))
>>> input_b = next(iter(dataloader))

>>> with accelerator.no_sync():
...     outputs = model(input_a)
...     loss = loss_func(outputs)
...     accelerator.backward(loss)
...     # No synchronization across processes, only accumulate gradients
>>> outputs = model(input_b)
>>> accelerator.backward(loss)
>>> # Synchronization across all processes
>>> optimizer.step()
>>> optimizer.zero_grad()

on_last_process

< >

( function: Callable[..., Any] )

Parameters

  • function (Callable) — The function to decorate.

A decorator that will run the decorated function on the last process only. Can also be called using the PartialState class.

Example:

# Assume we have 4 processes.
from accelerate import Accelerator

accelerator = Accelerator()


@accelerator.on_last_process
def print_something():
    print(f"Printed on process {accelerator.process_index}")


print_something()
"Printed on process 3"

on_local_main_process

< >

( function: Callable[..., Any] = None )

Parameters

  • function (Callable) — The function to decorate.

A decorator that will run the decorated function on the local main process only. Can also be called using the PartialState class.

Example:

# Assume we have 2 servers with 4 processes each.
from accelerate import Accelerator

accelerator = Accelerator()


@accelerator.on_local_main_process
def print_something():
    print("This will be printed by process 0 only on each server.")


print_something()
# On server 1:
"This will be printed by process 0 only"
# On server 2:
"This will be printed by process 0 only"

on_local_process

< >

( function: Callable[..., Any] = None local_process_index: int = None )

Parameters

  • function (Callable, optional) — The function to decorate.
  • local_process_index (int, optional) — The index of the local process on which to run the function.

A decorator that will run the decorated function on a given local process index only. Can also be called using the PartialState class.

Example:

# Assume we have 2 servers with 4 processes each.
from accelerate import Accelerator

accelerator = Accelerator()


@accelerator.on_local_process(local_process_index=2)
def print_something():
    print(f"Printed on process {accelerator.local_process_index}")


print_something()
# On server 1:
"Printed on process 2"
# On server 2:
"Printed on process 2"

on_main_process

< >

( function: Callable[..., Any] = None )

Parameters

  • function (Callable) — The function to decorate.

A decorator that will run the decorated function on the main process only. Can also be called using the PartialState class.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()


>>> @accelerator.on_main_process
... def print_something():
...     print("This will be printed by process 0 only.")


>>> print_something()
"This will be printed by process 0 only"

on_process

< >

( function: Callable[..., Any] = None process_index: int = None )

Parameters

  • function (Callable, optional) — The function to decorate.
  • process_index (int, optional) — The index of the process on which to run the function.

A decorator that will run the decorated function on a given process index only. Can also be called using the PartialState class.

Example:

# Assume we have 4 processes.
from accelerate import Accelerator

accelerator = Accelerator()


@accelerator.on_process(process_index=2)
def print_something():
    print(f"Printed on process {accelerator.process_index}")


print_something()
"Printed on process 2"

pad_across_processes

< >

( tensor dim = 0 pad_index = 0 pad_first = False ) → torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor

Parameters

  • tensor (nested list/tuple/dictionary of torch.Tensor) — The data to gather.
  • dim (int, optional, defaults to 0) — The dimension on which to pad.
  • pad_index (int, optional, defaults to 0) — The value with which to pad.
  • pad_first (bool, optional, defaults to False) — Whether to pad at the beginning or the end.

Returns

torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor

The padded tensor(s).

Recursively pad the tensors in a nested list/tuple/dictionary of tensors from all devices to the same size so they can safely be gathered.

Example:

>>> # Assuming two processes, with the first processes having a tensor of size 1 and the second of size 2
>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> process_tensor = torch.arange(accelerator.process_index + 1).to(accelerator.device)
>>> padded_tensor = accelerator.pad_across_processes(process_tensor)
>>> padded_tensor.shape
torch.Size([2])

prepare

< >

( *args device_placement = None )

Parameters

  • *args (list of objects) — Any of the following type of objects:

    • torch.utils.data.DataLoader: PyTorch Dataloader
    • torch.nn.Module: PyTorch Module
    • torch.optim.Optimizer: PyTorch Optimizer
    • torch.optim.lr_scheduler.LRScheduler: PyTorch LR Scheduler
  • device_placement (list[bool], optional) — Used to customize whether automatic device placement should be performed for each object passed. Needs to be a list of the same length as args. Not compatible with DeepSpeed or FSDP.

Prepare all objects passed in args for distributed training and mixed precision, then return them in the same order.

You don’t need to prepare a model if you only use it for inference without any kind of mixed precision

Examples:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> # Assume a model, optimizer, data_loader and scheduler are defined
>>> model, optimizer, data_loader, scheduler = accelerator.prepare(model, optimizer, data_loader, scheduler)
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> # Assume a model, optimizer, data_loader and scheduler are defined
>>> device_placement = [True, True, False, False]
>>> # Will place the first to items passed in automatically to the right device but not the last two.
>>> model, optimizer, data_loader, scheduler = accelerator.prepare(
...     model, optimizer, data_loader, scheduler, device_placement=device_placement
... )

prepare_data_loader

< >

( data_loader: torch.utils.data.DataLoader device_placement = None slice_fn_for_dispatch = None )

Parameters

  • data_loader (torch.utils.data.DataLoader) — A vanilla PyTorch DataLoader to prepare
  • device_placement (bool, optional) — Whether or not to place the batches on the proper device in the prepared dataloader. Will default to self.device_placement.
  • slice_fn_for_dispatch (Callable, optional) -- If passed, this function will be used to slice tensors across num_processes. Will default to slice_tensors(). This argument is used only when dispatch_batchesis set toTrue` and will be ignored otherwise.

Prepares a PyTorch DataLoader for training in any distributed setup. It is recommended to use Accelerator.prepare() instead.

Example:

>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> data_loader = torch.utils.data.DataLoader(...)
>>> data_loader = accelerator.prepare_data_loader(data_loader, device_placement=True)

prepare_model

< >

( model: torch.nn.Module device_placement: bool = None evaluation_mode: bool = False )

Parameters

  • model (torch.nn.Module) — A PyTorch model to prepare. You don’t need to prepare a model if it is used only for inference without any kind of mixed precision
  • device_placement (bool, optional) — Whether or not to place the model on the proper device. Will default to self.device_placement.
  • evaluation_mode (bool, optional, defaults to False) — Whether or not to set the model for evaluation only, by just applying mixed precision and torch.compile (if configured in the Accelerator object).

Prepares a PyTorch model for training in any distributed setup. It is recommended to use Accelerator.prepare() instead.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> # Assume a model is defined
>>> model = accelerator.prepare_model(model)

prepare_optimizer

< >

( optimizer: torch.optim.Optimizer device_placement = None )

Parameters

  • optimizer (torch.optim.Optimizer) — A vanilla PyTorch optimizer to prepare
  • device_placement (bool, optional) — Whether or not to place the optimizer on the proper device. Will default to self.device_placement.

Prepares a PyTorch Optimizer for training in any distributed setup. It is recommended to use Accelerator.prepare() instead.

Example:

>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> optimizer = torch.optim.Adam(...)
>>> optimizer = accelerator.prepare_optimizer(optimizer, device_placement=True)

prepare_scheduler

< >

( scheduler: LRScheduler )

Parameters

  • scheduler (torch.optim.lr_scheduler.LRScheduler) — A vanilla PyTorch scheduler to prepare

Prepares a PyTorch Scheduler for training in any distributed setup. It is recommended to use Accelerator.prepare() instead.

Example:

>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> optimizer = torch.optim.Adam(...)
>>> scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, ...)
>>> scheduler = accelerator.prepare_scheduler(scheduler)

print

< >

( *args **kwargs )

Drop in replacement of print() to only print once per server.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> accelerator.print("Hello world!")

reduce

< >

( tensor reduction = 'sum' scale = 1.0 ) → torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor

Parameters

  • tensor (torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor) — The tensors to reduce across all processes.
  • reduction (str, optional, defaults to “sum”) — A reduction type, can be one of ‘sum’, ‘mean’, or ‘none’. If ‘none’, will not perform any operation.
  • scale (float, optional, defaults to 1.0) — A default scaling value to be applied after the reduce, only valied on XLA.

Returns

torch.Tensor, or a nested tuple/list/dictionary of torch.Tensor

The reduced tensor(s).

Reduce the values in tensor across all processes based on reduction.

Note: All processes get the reduced value.

Example:

>>> # Assuming two processes
>>> import torch
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> process_tensor = torch.arange(accelerator.num_processes) + 1 + (2 * accelerator.process_index)
>>> process_tensor = process_tensor.to(accelerator.device)
>>> reduced_tensor = accelerator.reduce(process_tensor, reduction="sum")
>>> reduced_tensor
tensor([4, 6])

register_for_checkpointing

< >

( *objects )

Makes note of objects and will save or load them in during save_state or load_state.

These should be utilized when the state is being loaded or saved in the same script. It is not designed to be used in different scripts.

Every object must have a load_state_dict and state_dict function to be stored.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> # Assume `CustomObject` has a `state_dict` and `load_state_dict` function.
>>> obj = CustomObject()
>>> accelerator.register_for_checkpointing(obj)
>>> accelerator.save_state("checkpoint.pt")

register_load_state_pre_hook

< >

( hook: Callable[(Ellipsis, None)] ) → torch.utils.hooks.RemovableHandle

Parameters

Returns

torch.utils.hooks.RemovableHandle

a handle that can be used to remove the added hook by calling handle.remove()

Registers a pre hook to be run before load_checkpoint is called in Accelerator.load_state().

The hook should have the following signature:

hook(models: list[torch.nn.Module], input_dir: str) -> None

The models argument are the models as saved in the accelerator state under accelerator._models, and the input_dir argument is the input_dir argument passed to Accelerator.load_state().

Should only be used in conjunction with Accelerator.register_save_state_pre_hook(). Can be useful to load configurations in addition to model weights. Can also be used to overwrite model loading with a customized method. In this case, make sure to remove already loaded models from the models list.

register_save_state_pre_hook

< >

( hook: Callable[(Ellipsis, None)] ) → torch.utils.hooks.RemovableHandle

Parameters

Returns

torch.utils.hooks.RemovableHandle

a handle that can be used to remove the added hook by calling handle.remove()

Registers a pre hook to be run before save_checkpoint is called in Accelerator.save_state().

The hook should have the following signature:

hook(models: list[torch.nn.Module], weights: list[dict[str, torch.Tensor]], input_dir: str) -> None

The models argument are the models as saved in the accelerator state under accelerator._models, weigths argument are the state dicts of the models, and the input_dir argument is the input_dir argument passed to Accelerator.load_state().

Should only be used in conjunction with Accelerator.register_load_state_pre_hook(). Can be useful to save configurations in addition to model weights. Can also be used to overwrite model saving with a customized method. In this case, make sure to remove already loaded weights from the weights list.

save

< >

( obj f safe_serialization = False )

Parameters

  • obj (object) — The object to save.
  • f (str or os.PathLike) — Where to save the content of obj.
  • safe_serialization (bool, optional, defaults to False) — Whether to save obj using safetensors

Save the object passed to disk once per machine. Use in place of torch.save.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> arr = [0, 1, 2, 3]
>>> accelerator.save(arr, "array.pkl")

save_model

< >

( model: torch.nn.Module save_directory: Union[str, os.PathLike] max_shard_size: Union[int, str] = '10GB' safe_serialization: bool = False )

Parameters

  • save_directory (str or os.PathLike) — Directory to which to save. Will be created if it doesn’t exist.
  • max_shard_size (int or str, optional, defaults to "10GB") — The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size lower than this size. If expressed as a string, needs to be digits followed by a unit (like "5MB").

    If a single weight of the model is bigger than max_shard_size, it will be in its own checkpoint shard which will be bigger than max_shard_size.
  • safe_serialization (bool, optional, defaults to False) — Whether to save the model using safetensors or the traditional PyTorch way (that uses pickle).

Save a model so that it can be re-loaded using load_checkpoint_in_model

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> model = ...
>>> accelerator.save_model(model, save_directory)

save_state

< >

( output_dir: str = None **save_model_func_kwargs )

Parameters

  • output_dir (str or os.PathLike) — The name of the folder to save all relevant weights and states.
  • save_model_func_kwargs (dict, optional) — Additional keyword arguments for saving model which can be passed to the underlying save function, such as optional arguments for DeepSpeed’s save_checkpoint function.

Saves the current states of the model, optimizer, scaler, RNG generators, and registered objects to a folder.

If a ProjectConfiguration was passed to the Accelerator object with automatic_checkpoint_naming enabled then checkpoints will be saved to self.project_dir/checkpoints. If the number of current saves is greater than total_limit then the oldest save is deleted. Each checkpoint is saved in seperate folders named checkpoint_<iteration>.

Otherwise they are just saved to output_dir.

Should only be used when wanting to save a checkpoint during training and restoring the state in the same environment.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> model, optimizer, lr_scheduler = ...
>>> model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
>>> accelerator.save_state(output_dir="my_checkpoint")

set_trigger

< >

( )

Sets the internal trigger tensor to 1 on the current process. A latter check should follow using this which will check across all processes.

Note: Does not require wait_for_everyone()

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> # Assume later in the training script
>>> # `should_do_breakpoint` is a custom function to monitor when to break,
>>> # e.g. when the loss is NaN
>>> if should_do_breakpoint(loss):
...     accelerator.set_trigger()
>>> # Assume later in the training script
>>> if accelerator.check_breakpoint():
...     break

skip_first_batches

< >

( dataloader num_batches: int = 0 )

Parameters

  • dataloader (torch.utils.data.DataLoader) — The data loader in which to skip batches.
  • num_batches (int, optional, defaults to 0) — The number of batches to skip

Creates a new torch.utils.data.DataLoader that will efficiently skip the first num_batches.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
>>> skipped_dataloader = accelerator.skip_first_batches(dataloader, num_batches=2)
>>> # for the first epoch only
>>> for input, target in skipped_dataloader:
...     optimizer.zero_grad()
...     output = model(input)
...     loss = loss_func(output, target)
...     accelerator.backward(loss)
...     optimizer.step()

>>> # subsequent epochs
>>> for input, target in dataloader:
...     optimizer.zero_grad()
...     ...

split_between_processes

< >

( inputs: list | tuple | dict | torch.Tensor apply_padding: bool = False )

Parameters

  • inputs (list, tuple, torch.Tensor, or dict of list/tuple/torch.Tensor) — The input to split between processes.
  • apply_padding (bool, optional, defaults to False) — Whether to apply padding by repeating the last element of the input so that all processes have the same number of elements. Useful when trying to perform actions such as Accelerator.gather() on the outputs or passing in less inputs than there are processes. If so, just remember to drop the padded elements afterwards.

Splits input between self.num_processes quickly and can be then used on that process. Useful when doing distributed inference, such as with different prompts.

Note that when using a dict, all keys need to have the same number of elements.

Example:

# Assume there are two processes
from accelerate import Accelerator

accelerator = Accelerator()
with accelerator.split_between_processes(["A", "B", "C"]) as inputs:
    print(inputs)
# Process 0
["A", "B"]
# Process 1
["C"]

with accelerator.split_between_processes(["A", "B", "C"], apply_padding=True) as inputs:
    print(inputs)
# Process 0
["A", "B"]
# Process 1
["C", "C"]

trigger_sync_in_backward

< >

( model )

Parameters

  • model (torch.nn.Module) — The model for which to trigger the gradient synchronization.

Trigger the sync of the gradients in the next backward pass of the model after multiple forward passes under Accelerator.no_sync (only applicable in multi-GPU scenarios).

If the script is not launched in distributed mode, this context manager does nothing.

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> dataloader, model, optimizer = accelerator.prepare(dataloader, model, optimizer)

>>> with accelerator.no_sync():
...     loss_a = loss_func(model(input_a))  # first forward pass
...     loss_b = loss_func(model(input_b))  # second forward pass
>>> accelerator.backward(loss_a)  # No synchronization across processes, only accumulate gradients
>>> with accelerator.trigger_sync_in_backward(model):
...     accelerator.backward(loss_b)  # Synchronization across all processes
>>> optimizer.step()
>>> optimizer.zero_grad()

unscale_gradients

< >

( optimizer = None )

Parameters

  • optimizer (torch.optim.Optimizer or list[torch.optim.Optimizer], optional) — The optimizer(s) for which to unscale gradients. If not set, will unscale gradients on all optimizers that were passed to prepare().

Unscale the gradients in mixed precision training with AMP. This is a noop in all other settings.

Likely should be called through Accelerator.clipgrad_norm() or Accelerator.clipgrad_value()

Example:

>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> model, optimizer = accelerator.prepare(model, optimizer)
>>> outputs = model(inputs)
>>> loss = loss_fn(outputs, labels)
>>> accelerator.backward(loss)
>>> accelerator.unscale_gradients(optimizer=optimizer)

unwrap_model

< >

( model keep_fp32_wrapper: bool = True ) → torch.nn.Module

Parameters

  • model (torch.nn.Module) — The model to unwrap.
  • keep_fp32_wrapper (bool, optional, defaults to True) — Whether to not remove the mixed precision hook if it was added.

Returns

torch.nn.Module

The unwrapped model.

Unwraps the model from the additional layer possible added by prepare(). Useful before saving the model.

Example:

>>> # Assuming two GPU processes
>>> from torch.nn.parallel import DistributedDataParallel
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> model = accelerator.prepare(MyModel())
>>> print(model.__class__.__name__)
DistributedDataParallel

>>> model = accelerator.unwrap_model(model)
>>> print(model.__class__.__name__)
MyModel

verify_device_map

< >

( model: torch.nn.Module )

Verifies that model has not been prepared with big model inference with a device-map resembling auto.

wait_for_everyone

< >

( )

Will stop the execution of the current process until every other process has reached that point (so this does nothing when the script is only run in one process). Useful to do before saving a model.

Example:

>>> # Assuming two GPU processes
>>> import time
>>> from accelerate import Accelerator

>>> accelerator = Accelerator()
>>> if accelerator.is_main_process:
...     time.sleep(2)
>>> else:
...     print("I'm waiting for the main process to finish its sleep...")
>>> accelerator.wait_for_everyone()
>>> # Should print on every process at the same time
>>> print("Everyone is here")
←TPU best practices Stateful configuration classes→