Datasets:

muellerzr
/

RAG-accelerate

Languages:

English

License:

Dataset card Files Files and versions Community

muellerzr HF staff commited on Jan 2

Commit

5887b84

•

1 Parent(s): 60de117

Spacing

Browse files

Files changed (20) hide show

README.md +9 -1
src/accelerator.py +1 -4
src/commands/config/config_utils.py +0 -1
src/commands/launch.py +0 -1
src/data_loader.py +0 -8
src/hooks.py +0 -7
src/local_sgd.py +0 -2
src/logging.py +0 -1
src/optimizer.py +0 -1
src/scheduler.py +0 -1
src/state.py +0 -4
src/tracking.py +0 -9
src/utils/bnb.py +0 -2
src/utils/dataclasses.py +0 -20
src/utils/deepspeed.py +0 -6
src/utils/launch.py +0 -1
src/utils/megatron_lm.py +0 -11
src/utils/modeling.py +0 -3
src/utils/offload.py +0 -2
src/utils/operations.py +0 -12

README.md CHANGED Viewed

@@ -37,4 +37,12 @@ Using `regex` in VSCODE, use the following replacement:
 <!--Copyright(.*\n)+-->
 ```
-Then remove all import statements (as we only care about the content).

 <!--Copyright(.*\n)+-->
 ```
+In the source:
+```regex
+    """
+```
+Then remove all import statements (as we only care about the content).
+**WARNING**: It is known that this will seperate out the `_inner()` in the source code and use it as a seperate function losing the context. Trying out with this issue for now.

src/accelerator.py CHANGED Viewed

@@ -1,3 +1,4 @@
 logger = get_logger(__name__)
@@ -95,7 +96,6 @@ class Accelerator:
         - **sync_gradients** (`bool`) -- Whether the gradients are currently being synced across all processes.
         - **use_distributed** (`bool`) -- Whether the current configuration is for distributed training.
     """
     def __init__(
         self,
         device_placement: bool = True,
@@ -2010,7 +2010,6 @@ class Accelerator:
         9
         ```
         """
         try:
             recursively_apply(lambda x: x, input_data, error_on_other_type=True)
             all_tensors = True
@@ -2373,7 +2372,6 @@ class Accelerator:
         >>> accelerator.save_model(model, save_directory)
         ```
         """
         if os.path.isfile(save_directory):
             logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
             return
@@ -2865,7 +2863,6 @@ class Accelerator:
         >>> state_dict = accelerator.get_state_dict(net)
         ```
         """
         if self.distributed_type == DistributedType.DEEPSPEED:
             if self.deepspeed_config["zero_optimization"]["stage"] == 3:
                 if model.zero_gather_16bit_weights_on_model_save():

 logger = get_logger(__name__)
         - **sync_gradients** (`bool`) -- Whether the gradients are currently being synced across all processes.
         - **use_distributed** (`bool`) -- Whether the current configuration is for distributed training.
     """
     def __init__(
         self,
         device_placement: bool = True,
         9
         ```
         """
         try:
             recursively_apply(lambda x: x, input_data, error_on_other_type=True)
             all_tensors = True
         >>> accelerator.save_model(model, save_directory)
         ```
         """
         if os.path.isfile(save_directory):
             logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
             return
         >>> state_dict = accelerator.get_state_dict(net)
         ```
         """
         if self.distributed_type == DistributedType.DEEPSPEED:
             if self.deepspeed_config["zero_optimization"]["stage"] == 3:
                 if model.zero_gather_16bit_weights_on_model_save():

src/commands/config/config_utils.py CHANGED Viewed

@@ -67,7 +67,6 @@ class SubcommandHelpFormatter(argparse.RawDescriptionHelpFormatter):
     """
     A custom formatter that will remove the usage line from the help message for subcommands.
     """
     def _format_usage(self, usage, actions, groups, prefix):
         usage = super()._format_usage(usage, actions, groups, prefix)
         usage = usage.replace("<command> [<args>] ", "")

     """
     A custom formatter that will remove the usage line from the help message for subcommands.
     """
     def _format_usage(self, usage, actions, groups, prefix):
         usage = super()._format_usage(usage, actions, groups, prefix)
         usage = usage.replace("<command> [<args>] ", "")

src/commands/launch.py CHANGED Viewed

@@ -22,7 +22,6 @@ class _CustomHelpAction(argparse._HelpAction):
     called. This is useful for the case where the user is using a specific platform and only wants to see the arguments
     for that platform.
     """
     def __call__(self, parser, namespace, values, option_string=None):
         if "accelerate" in sys.argv[0] and "launch" in sys.argv[1:]:
             args = sys.argv[2:]

     called. This is useful for the case where the user is using a specific platform and only wants to see the arguments
     for that platform.
     """
     def __call__(self, parser, namespace, values, option_string=None):
         if "accelerate" in sys.argv[0] and "launch" in sys.argv[1:]:
             args = sys.argv[2:]

src/data_loader.py CHANGED Viewed

@@ -36,7 +36,6 @@ class SeedableRandomSampler(RandomSampler):
     If a custom `generator` is passed, it will rely on its initial seed as well as the current iteration it is on
     (stored in `self.epoch`).
     """
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.epoch = 0
@@ -246,7 +245,6 @@ class IterableDatasetShard(IterableDataset):
             - the shard on process 0 to yield `[0, 1, 4, 5]` and the sampler on process 1 to yield `[2, 3, 6, 7]` if
               this argument is set to `True`.
     """
     def __init__(
         self,
         dataset: IterableDataset,
@@ -326,7 +324,6 @@ class DataLoaderStateMixin:
           batch size
     """
     def __init_subclass__(cls, **kwargs):
         cls.end_of_dataloader = False
         cls.remainder = -1
@@ -381,7 +378,6 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
         - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
     """
     def __init__(
         self,
         dataset,
@@ -481,7 +477,6 @@ if is_tpu_available(check_device=False):
         - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
         """
         def __init__(self, dataloader: DataLoaderShard, device: torch.device):
             super().__init__(dataloader, device)
             self._rng_types = self._loader.rng_types
@@ -530,7 +525,6 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
         - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
     """
     def __init__(
         self, dataset, split_batches: bool = False, skip_batches=0, _drop_last: bool = False, slice_fn=None, **kwargs
     ):
@@ -907,7 +901,6 @@ class SkipBatchSampler(BatchSampler):
     """
     A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
     """
     def __init__(self, batch_sampler, skip_batches=0):
         self.batch_sampler = batch_sampler
         self.skip_batches = skip_batches
@@ -937,7 +930,6 @@ class SkipDataLoader(DataLoader):
         kwargs:
             All other keyword arguments to pass to the regular `DataLoader` initialization.
     """
     def __init__(self, dataset, skip_batches=0, **kwargs):
         super().__init__(dataset, **kwargs)
         self.skip_batches = skip_batches

     If a custom `generator` is passed, it will rely on its initial seed as well as the current iteration it is on
     (stored in `self.epoch`).
     """
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.epoch = 0
             - the shard on process 0 to yield `[0, 1, 4, 5]` and the sampler on process 1 to yield `[2, 3, 6, 7]` if
               this argument is set to `True`.
     """
     def __init__(
         self,
         dataset: IterableDataset,
           batch size
     """
     def __init_subclass__(cls, **kwargs):
         cls.end_of_dataloader = False
         cls.remainder = -1
         - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
     """
     def __init__(
         self,
         dataset,
         - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
         """
         def __init__(self, dataloader: DataLoaderShard, device: torch.device):
             super().__init__(dataloader, device)
             self._rng_types = self._loader.rng_types
         - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
     """
     def __init__(
         self, dataset, split_batches: bool = False, skip_batches=0, _drop_last: bool = False, slice_fn=None, **kwargs
     ):
     """
     A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
     """
     def __init__(self, batch_sampler, skip_batches=0):
         self.batch_sampler = batch_sampler
         self.skip_batches = skip_batches
         kwargs:
             All other keyword arguments to pass to the regular `DataLoader` initialization.
     """
     def __init__(self, dataset, skip_batches=0, **kwargs):
         super().__init__(dataset, **kwargs)
         self.skip_batches = skip_batches

src/hooks.py CHANGED Viewed

@@ -7,7 +7,6 @@ class ModelHook:
     - **no_grad** (`bool`, *optional*, defaults to `False`) -- Whether or not to execute the actual forward pass under
       the `torch.no_grad()` context manager.
     """
     no_grad = False
     def init_hook(self, module):
@@ -60,7 +59,6 @@ class SequentialHook(ModelHook):
     """
     A hook that can contain several hooks and iterates through them at each event.
     """
     def __init__(self, *hooks):
         self.hooks = hooks
@@ -109,7 +107,6 @@ def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False)
         `torch.nn.Module`: The same module, with the hook attached (the module is modified in place, so the result can
         be discarded).
     """
     if append and (getattr(module, "_hf_hook", None) is not None):
         old_hook = module._hf_hook
         remove_hook_from_module(module)
@@ -151,7 +148,6 @@ def remove_hook_from_module(module: nn.Module, recurse=False):
         `torch.nn.Module`: The same module, with the hook detached (the module is modified in place, so the result can
         be discarded).
     """
     if hasattr(module, "_hf_hook"):
         module._hf_hook.detach_hook(module)
         delattr(module, "_hf_hook")
@@ -186,7 +182,6 @@ class AlignDevicesHook(ModelHook):
         place_submodules (`bool`, *optional*, defaults to `False`):
             Whether to place the submodules on `execution_device` during the `init_hook` event.
     """
     def __init__(
         self,
         execution_device: Optional[Union[int, str, torch.device]] = None,
@@ -539,7 +534,6 @@ class CpuOffload(ModelHook):
             passed, its offload method will be called just before the forward of the model to which this hook is
             attached.
     """
     def __init__(
         self,
         execution_device: Optional[Union[str, int, torch.device]] = None,
@@ -564,7 +558,6 @@ class UserCpuOffloadHook:
     A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
     or remove it entirely.
     """
     def __init__(self, model, hook):
         self.model = model
         self.hook = hook

     - **no_grad** (`bool`, *optional*, defaults to `False`) -- Whether or not to execute the actual forward pass under
       the `torch.no_grad()` context manager.
     """
     no_grad = False
     def init_hook(self, module):
     """
     A hook that can contain several hooks and iterates through them at each event.
     """
     def __init__(self, *hooks):
         self.hooks = hooks
         `torch.nn.Module`: The same module, with the hook attached (the module is modified in place, so the result can
         be discarded).
     """
     if append and (getattr(module, "_hf_hook", None) is not None):
         old_hook = module._hf_hook
         remove_hook_from_module(module)
         `torch.nn.Module`: The same module, with the hook detached (the module is modified in place, so the result can
         be discarded).
     """
     if hasattr(module, "_hf_hook"):
         module._hf_hook.detach_hook(module)
         delattr(module, "_hf_hook")
         place_submodules (`bool`, *optional*, defaults to `False`):
             Whether to place the submodules on `execution_device` during the `init_hook` event.
     """
     def __init__(
         self,
         execution_device: Optional[Union[int, str, torch.device]] = None,
             passed, its offload method will be called just before the forward of the model to which this hook is
             attached.
     """
     def __init__(
         self,
         execution_device: Optional[Union[str, int, torch.device]] = None,
     A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
     or remove it entirely.
     """
     def __init__(self, model, hook):
         self.model = model
         self.hook = hook

src/local_sgd.py CHANGED Viewed

@@ -19,7 +19,6 @@ class LocalSGD:
     Learning Representations. No. CONF. 2019.](https://arxiv.org/abs/1805.09767)
     """
     def __enter__(self):
         if self.enabled:
             self.model_sync_obj = self.model.no_sync()
@@ -75,7 +74,6 @@ class LocalSGD:
         """
         Synchronize + Average model parameters across all GPUs
         """
         self.accelerator.wait_for_everyone()
         with self.accelerator.autocast():
             for param in self.model.parameters():

     Learning Representations. No. CONF. 2019.](https://arxiv.org/abs/1805.09767)
     """
     def __enter__(self):
         if self.enabled:
             self.model_sync_obj = self.model.no_sync()
         """
         Synchronize + Average model parameters across all GPUs
         """
         self.accelerator.wait_for_everyone()
         with self.accelerator.autocast():
             for param in self.model.parameters():

src/logging.py CHANGED Viewed

@@ -7,7 +7,6 @@ class MultiProcessAdapter(logging.LoggerAdapter):
     Does not require an `Accelerator` object to be created first.
     """
     @staticmethod
     def _should_log(main_process_only):
         "Check if log should be performed"

     Does not require an `Accelerator` object to be created first.
     """
     @staticmethod
     def _should_log(main_process_only):
         "Check if log should be performed"

src/optimizer.py CHANGED Viewed

@@ -24,7 +24,6 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
         scaler (`torch.cuda.amp.grad_scaler.GradScaler`, *optional*):
             The scaler to use in the step function if training with mixed precision.
     """
     def __init__(self, optimizer, device_placement=True, scaler=None):
         self.optimizer = optimizer
         self.scaler = scaler

         scaler (`torch.cuda.amp.grad_scaler.GradScaler`, *optional*):
             The scaler to use in the step function if training with mixed precision.
     """
     def __init__(self, optimizer, device_placement=True, scaler=None):
         self.optimizer = optimizer
         self.scaler = scaler

src/scheduler.py CHANGED Viewed

@@ -23,7 +23,6 @@ class AcceleratedScheduler:
             regardless of the number of processes) or create batches on each process (so batch size is the original
             batch size multiplied by the number of processes).
     """
     def __init__(self, scheduler, optimizers, step_with_optimizer: bool = True, split_batches: bool = False):
         self.scheduler = scheduler
         self.optimizers = optimizers if isinstance(optimizers, (list, tuple)) else [optimizers]

             regardless of the number of processes) or create batches on each process (so batch size is the original
             batch size multiplied by the number of processes).
     """
     def __init__(self, scheduler, optimizers, step_with_optimizer: bool = True, split_batches: bool = False):
         self.scheduler = scheduler
         self.optimizers = optimizers if isinstance(optimizers, (list, tuple)) else [optimizers]

src/state.py CHANGED Viewed

@@ -30,7 +30,6 @@ class ThreadLocalSharedDict(threading.local):
     See https://github.com/pytorch/xla/blob/r2.0/docs/pjrt.md#multithreading-on-tpu-v2v3
     """
     def __init__(self, thread_local: bool = False):
         self._storage = {}
@@ -67,7 +66,6 @@ class PartialState:
         - **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
         - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
     """
     _shared_state = SharedDict()
     def __init__(self, cpu: bool = False, **kwargs):
@@ -684,7 +682,6 @@ class AcceleratorState:
         - **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
         - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
     """
     _shared_state = SharedDict()
     def __init__(
@@ -946,7 +943,6 @@ class GradientState:
         - **sync_with_dataloader** (`bool`) -- Whether the gradients should be synced at the end of the dataloader
             iteration and the number of total steps reset
     """
     _shared_state = SharedDict()
     def __init__(self, gradient_accumulation_plugin: Optional[GradientAccumulationPlugin] = None):

     See https://github.com/pytorch/xla/blob/r2.0/docs/pjrt.md#multithreading-on-tpu-v2v3
     """
     def __init__(self, thread_local: bool = False):
         self._storage = {}
         - **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
         - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
     """
     _shared_state = SharedDict()
     def __init__(self, cpu: bool = False, **kwargs):
         - **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
         - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
     """
     _shared_state = SharedDict()
     def __init__(
         - **sync_with_dataloader** (`bool`) -- Whether the gradients should be synced at the end of the dataloader
             iteration and the number of total steps reset
     """
     _shared_state = SharedDict()
     def __init__(self, gradient_accumulation_plugin: Optional[GradientAccumulationPlugin] = None):

src/tracking.py CHANGED Viewed

@@ -37,7 +37,6 @@ def on_main_process(function):
     Checks at function execution rather than initialization time, not triggering the initialization of the
     `PartialState`.
     """
     @wraps(function)
     def execute_on_main_process(self, *args, **kwargs):
         if getattr(self, "main_process_only", False):
@@ -69,7 +68,6 @@ class GeneralTracker:
     Implementations can also include a `main_process_only` (`bool`) attribute to toggle if relevent logging, init, and
     other functions should occur on the main process or across all processes (by default will use `True`)
     """
     main_process_only = True
     def __init__(self, _blank=False):
@@ -139,7 +137,6 @@ class TensorBoardTracker(GeneralTracker):
         kwargs:
             Additional key word arguments passed along to the `tensorboard.SummaryWriter.__init__` method.
     """
     name = "tensorboard"
     requires_logging_directory = True
@@ -248,7 +245,6 @@ class WandBTracker(GeneralTracker):
         kwargs:
             Additional key word arguments passed along to the `wandb.init` method.
     """
     name = "wandb"
     requires_logging_directory = False
     main_process_only = False
@@ -373,7 +369,6 @@ class CometMLTracker(GeneralTracker):
         kwargs:
             Additional key word arguments passed along to the `Experiment.__init__` method.
     """
     name = "comet_ml"
     requires_logging_directory = False
@@ -452,7 +447,6 @@ class AimTracker(GeneralTracker):
         kwargs:
             Additional key word arguments passed along to the `Run.__init__` method.
     """
     name = "aim"
     requires_logging_directory = True
@@ -568,7 +562,6 @@ class MLflowTracker(GeneralTracker):
             description is set on the resumed run. If a new run is being created, the description is set on the new
             run.
     """
     name = "mlflow"
     requires_logging_directory = False
@@ -697,7 +690,6 @@ class ClearMLTracker(GeneralTracker):
         kwargs:
             Kwargs passed along to the `Task.__init__` method.
     """
     name = "clearml"
     requires_logging_directory = False
@@ -857,7 +849,6 @@ class DVCLiveTracker(GeneralTracker):
     accelerator.init_trackers(project_name="my_project", init_kwargs={"dvclive": {"dir": "my_directory"}})
     ```
     """
     name = "dvclive"
     requires_logging_directory = False

     Checks at function execution rather than initialization time, not triggering the initialization of the
     `PartialState`.
     """
     @wraps(function)
     def execute_on_main_process(self, *args, **kwargs):
         if getattr(self, "main_process_only", False):
     Implementations can also include a `main_process_only` (`bool`) attribute to toggle if relevent logging, init, and
     other functions should occur on the main process or across all processes (by default will use `True`)
     """
     main_process_only = True
     def __init__(self, _blank=False):
         kwargs:
             Additional key word arguments passed along to the `tensorboard.SummaryWriter.__init__` method.
     """
     name = "tensorboard"
     requires_logging_directory = True
         kwargs:
             Additional key word arguments passed along to the `wandb.init` method.
     """
     name = "wandb"
     requires_logging_directory = False
     main_process_only = False
         kwargs:
             Additional key word arguments passed along to the `Experiment.__init__` method.
     """
     name = "comet_ml"
     requires_logging_directory = False
         kwargs:
             Additional key word arguments passed along to the `Run.__init__` method.
     """
     name = "aim"
     requires_logging_directory = True
             description is set on the resumed run. If a new run is being created, the description is set on the new
             run.
     """
     name = "mlflow"
     requires_logging_directory = False
         kwargs:
             Kwargs passed along to the `Task.__init__` method.
     """
     name = "clearml"
     requires_logging_directory = False
     accelerator.init_trackers(project_name="my_project", init_kwargs={"dvclive": {"dir": "my_directory"}})
     ```
     """
     name = "dvclive"
     requires_logging_directory = False

src/utils/bnb.py CHANGED Viewed

@@ -44,7 +44,6 @@ def load_and_quantize_model(
     Returns:
         `torch.nn.Module`: The quantized model
     """
     load_in_4bit = bnb_quantization_config.load_in_4bit
     load_in_8bit = bnb_quantization_config.load_in_8bit
@@ -246,7 +245,6 @@ def replace_with_bnb_layers(model, bnb_quantization_config, modules_to_not_conve
             An array to track the current key of the recursion. This is used to check whether the current key (part of
             it) is not in the list of modules to not convert.
     """
     if modules_to_not_convert is None:
         modules_to_not_convert = []

     Returns:
         `torch.nn.Module`: The quantized model
     """
     load_in_4bit = bnb_quantization_config.load_in_4bit
     load_in_8bit = bnb_quantization_config.load_in_8bit
             An array to track the current key of the recursion. This is used to check whether the current key (part of
             it) is not in the list of modules to not convert.
     """
     if modules_to_not_convert is None:
         modules_to_not_convert = []

src/utils/dataclasses.py CHANGED Viewed

@@ -5,7 +5,6 @@ class KwargsHandler:
     """
     Internal mixin that implements a `to_kwargs()` method for a dataclass.
     """
     def to_dict(self):
         return copy.deepcopy(self.__dict__)
@@ -39,7 +38,6 @@ class AutocastKwargs(KwargsHandler):
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     enabled: bool = True
     cache_enabled: bool = None
@@ -70,7 +68,6 @@ class DistributedDataParallelKwargs(KwargsHandler):
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     dim: int = 0
     broadcast_buffers: bool = True
     bucket_cap_mb: int = 25
@@ -103,7 +100,6 @@ class GradScalerKwargs(KwargsHandler):
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     init_scale: float = 65536.0
     growth_factor: float = 2.0
     backoff_factor: float = 0.5
@@ -128,7 +124,6 @@ class InitProcessGroupKwargs(KwargsHandler):
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     backend: Optional[str] = "nccl"
     init_method: Optional[str] = None
     timeout: timedelta = timedelta(seconds=1800)
@@ -197,7 +192,6 @@ class FP8RecipeKwargs(KwargsHandler):
                     are stored in FP8. If `fp8` is selected and deepspeed is enabled, will be used by default. (Not
                     available currently).
     """
     backend: Backend = "msamp"
     opt_level: OptLevel = "O2"
     margin: int = 0
@@ -260,7 +254,6 @@ class DistributedType(str, enum.Enum):
         - **DEEPSPEED** -- Using DeepSpeed.
         - **TPU** -- Distributed on TPUs.
     """
     # Subclassing str as well as Enum allows the `DistributedType` to be JSON-serializable out of the box.
     NO = "NO"
     MULTI_CPU = "MULTI_CPU"
@@ -283,7 +276,6 @@ class SageMakerDistributedType(str, enum.Enum):
         - **DATA_PARALLEL** -- using sagemaker distributed data parallelism.
         - **MODEL_PARALLEL** -- using sagemaker distributed model parallelism.
     """
     # Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
     NO = "NO"
     DATA_PARALLEL = "DATA_PARALLEL"
@@ -299,7 +291,6 @@ class ComputeEnvironment(str, enum.Enum):
         - **LOCAL_MACHINE** -- private/custom cluster hardware.
         - **AMAZON_SAGEMAKER** -- Amazon SageMaker as compute environment.
     """
     # Subclassing str as well as Enum allows the `ComputeEnvironment` to be JSON-serializable out of the box.
     LOCAL_MACHINE = "LOCAL_MACHINE"
     AMAZON_SAGEMAKER = "AMAZON_SAGEMAKER"
@@ -336,7 +327,6 @@ class DynamoBackend(str, BaseEnum):
         - **TVM** -- Uses Apach TVM for inference optimizations. [Read more](https://tvm.apache.org/)
     """
     # Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
     NO = "NO"
     EAGER = "EAGER"
@@ -364,7 +354,6 @@ class LoggerType(BaseEnum):
         - **COMETML** -- comet_ml as an experiment tracker
         - **DVCLIVE** -- dvclive as an experiment tracker
     """
     ALL = "all"
     AIM = "aim"
     TENSORBOARD = "tensorboard"
@@ -384,7 +373,6 @@ class PrecisionType(BaseEnum):
         - **FP16** -- using half precision
         - **BF16** -- using brain floating point precision
     """
     NO = "no"
     FP8 = "fp8"
     FP16 = "fp16"
@@ -404,7 +392,6 @@ class CustomDtype(enum.Enum):
     r"""
     An enum that contains multiple custom dtypes that can be used for `infer_auto_device_map`.
     """
     FP8 = "fp8"
     INT4 = "int4"
@@ -423,7 +410,6 @@ class ProjectConfiguration:
     """
     Configuration for the Accelerator object based on inner-project needs.
     """
     project_dir: str = field(default=None, metadata={"help": "A path to a directory for storing data."})
     logging_dir: str = field(
         default=None,
@@ -471,7 +457,6 @@ class GradientAccumulationPlugin(KwargsHandler):
     """
     A plugin to configure gradient accumulation behavior.
     """
     num_steps: int = field(default=None, metadata={"help": "The number of steps to accumulate gradients for."})
     adjust_scheduler: bool = field(
         default=True,
@@ -492,7 +477,6 @@ class TorchDynamoPlugin(KwargsHandler):
     """
     This plugin is used to compile a model with PyTorch 2.0
     """
     backend: DynamoBackend = field(
         default=None,
         metadata={"help": f"Possible options are {[b.value.lower() for b in DynamoBackend]}"},
@@ -528,7 +512,6 @@ class DeepSpeedPlugin:
     """
     This plugin is used to integrate DeepSpeed.
     """
     hf_ds_config: Any = field(
         default=None,
         metadata={
@@ -828,7 +811,6 @@ class FullyShardedDataParallelPlugin:
     """
     This plugin is used to enable fully sharded data parallelism.
     """
     sharding_strategy: "typing.Any" = field(
         default=None,
         metadata={
@@ -1062,7 +1044,6 @@ class MegatronLMPlugin:
     Plugin for Megatron-LM to enable tensor, pipeline, sequence and data parallelism. Also to enable selective
     activation recomputation and optimized fused kernels.
     """
     tp_degree: int = field(default=None, metadata={"help": "tensor parallelism degree."})
     pp_degree: int = field(default=None, metadata={"help": "pipeline parallelism degree."})
     num_micro_batches: int = field(default=None, metadata={"help": "number of micro-batches."})
@@ -1436,7 +1417,6 @@ class BnbQuantizationConfig:
     """
     A plugin to enable BitsAndBytes 4bit and 8bit quantization
     """
     load_in_8bit: bool = field(default=False, metadata={"help": "enable 8bit quantization."})
     llm_int8_threshold: float = field(

     """
     Internal mixin that implements a `to_kwargs()` method for a dataclass.
     """
     def to_dict(self):
         return copy.deepcopy(self.__dict__)
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     enabled: bool = True
     cache_enabled: bool = None
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     dim: int = 0
     broadcast_buffers: bool = True
     bucket_cap_mb: int = 25
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     init_scale: float = 65536.0
     growth_factor: float = 2.0
     backoff_factor: float = 0.5
     accelerator = Accelerator(kwargs_handlers=[kwargs])
     ```
     """
     backend: Optional[str] = "nccl"
     init_method: Optional[str] = None
     timeout: timedelta = timedelta(seconds=1800)
                     are stored in FP8. If `fp8` is selected and deepspeed is enabled, will be used by default. (Not
                     available currently).
     """
     backend: Backend = "msamp"
     opt_level: OptLevel = "O2"
     margin: int = 0
         - **DEEPSPEED** -- Using DeepSpeed.
         - **TPU** -- Distributed on TPUs.
     """
     # Subclassing str as well as Enum allows the `DistributedType` to be JSON-serializable out of the box.
     NO = "NO"
     MULTI_CPU = "MULTI_CPU"
         - **DATA_PARALLEL** -- using sagemaker distributed data parallelism.
         - **MODEL_PARALLEL** -- using sagemaker distributed model parallelism.
     """
     # Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
     NO = "NO"
     DATA_PARALLEL = "DATA_PARALLEL"
         - **LOCAL_MACHINE** -- private/custom cluster hardware.
         - **AMAZON_SAGEMAKER** -- Amazon SageMaker as compute environment.
     """
     # Subclassing str as well as Enum allows the `ComputeEnvironment` to be JSON-serializable out of the box.
     LOCAL_MACHINE = "LOCAL_MACHINE"
     AMAZON_SAGEMAKER = "AMAZON_SAGEMAKER"
         - **TVM** -- Uses Apach TVM for inference optimizations. [Read more](https://tvm.apache.org/)
     """
     # Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
     NO = "NO"
     EAGER = "EAGER"
         - **COMETML** -- comet_ml as an experiment tracker
         - **DVCLIVE** -- dvclive as an experiment tracker
     """
     ALL = "all"
     AIM = "aim"
     TENSORBOARD = "tensorboard"
         - **FP16** -- using half precision
         - **BF16** -- using brain floating point precision
     """
     NO = "no"
     FP8 = "fp8"
     FP16 = "fp16"
     r"""
     An enum that contains multiple custom dtypes that can be used for `infer_auto_device_map`.
     """
     FP8 = "fp8"
     INT4 = "int4"
     """
     Configuration for the Accelerator object based on inner-project needs.
     """
     project_dir: str = field(default=None, metadata={"help": "A path to a directory for storing data."})
     logging_dir: str = field(
         default=None,
     """
     A plugin to configure gradient accumulation behavior.
     """
     num_steps: int = field(default=None, metadata={"help": "The number of steps to accumulate gradients for."})
     adjust_scheduler: bool = field(
         default=True,
     """
     This plugin is used to compile a model with PyTorch 2.0
     """
     backend: DynamoBackend = field(
         default=None,
         metadata={"help": f"Possible options are {[b.value.lower() for b in DynamoBackend]}"},
     """
     This plugin is used to integrate DeepSpeed.
     """
     hf_ds_config: Any = field(
         default=None,
         metadata={
     """
     This plugin is used to enable fully sharded data parallelism.
     """
     sharding_strategy: "typing.Any" = field(
         default=None,
         metadata={
     Plugin for Megatron-LM to enable tensor, pipeline, sequence and data parallelism. Also to enable selective
     activation recomputation and optimized fused kernels.
     """
     tp_degree: int = field(default=None, metadata={"help": "tensor parallelism degree."})
     pp_degree: int = field(default=None, metadata={"help": "pipeline parallelism degree."})
     num_micro_batches: int = field(default=None, metadata={"help": "number of micro-batches."})
     """
     A plugin to enable BitsAndBytes 4bit and 8bit quantization
     """
     load_in_8bit: bool = field(default=False, metadata={"help": "enable 8bit quantization."})
     llm_int8_threshold: float = field(

src/utils/deepspeed.py CHANGED Viewed

@@ -14,7 +14,6 @@ class HfDeepSpeedConfig:
         config_file_or_dict (`Union[str, Dict]`): path to DeepSpeed config file or dict.
     """
     def __init__(self, config_file_or_dict):
         if isinstance(config_file_or_dict, dict):
             # Don't modify user's data should they want to reuse it (e.g. in tests), because once we
@@ -134,7 +133,6 @@ class DeepSpeedEngineWrapper:
     Args:
         engine (deepspeed.runtime.engine.DeepSpeedEngine): deepspeed engine to wrap
     """
     def __init__(self, engine):
         self.engine = engine
@@ -163,7 +161,6 @@ class DeepSpeedOptimizerWrapper(AcceleratedOptimizer):
         optimizer (`torch.optim.optimizer.Optimizer`):
             The optimizer to wrap.
     """
     def __init__(self, optimizer):
         super().__init__(optimizer, device_placement=False, scaler=None)
         self.__has_overflow__ = hasattr(self.optimizer, "overflow")
@@ -191,7 +188,6 @@ class DeepSpeedSchedulerWrapper(AcceleratedScheduler):
             The scheduler to wrap.
         optimizers (one or a list of `torch.optim.Optimizer`):
     """
     def __init__(self, scheduler, optimizers):
         super().__init__(scheduler, optimizers)
@@ -214,7 +210,6 @@ class DummyOptim:
         **kwargs:
             Other arguments.
     """
     def __init__(self, params, lr=0.001, weight_decay=0, **kwargs):
         self.params = params
         self.lr = lr
@@ -239,7 +234,6 @@ class DummyScheduler:
         **kwargs:
             Other arguments.
     """
     def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, lr_scheduler_callable=None, **kwargs):
         self.optimizer = optimizer
         self.total_num_steps = total_num_steps

         config_file_or_dict (`Union[str, Dict]`): path to DeepSpeed config file or dict.
     """
     def __init__(self, config_file_or_dict):
         if isinstance(config_file_or_dict, dict):
             # Don't modify user's data should they want to reuse it (e.g. in tests), because once we
     Args:
         engine (deepspeed.runtime.engine.DeepSpeedEngine): deepspeed engine to wrap
     """
     def __init__(self, engine):
         self.engine = engine
         optimizer (`torch.optim.optimizer.Optimizer`):
             The optimizer to wrap.
     """
     def __init__(self, optimizer):
         super().__init__(optimizer, device_placement=False, scaler=None)
         self.__has_overflow__ = hasattr(self.optimizer, "overflow")
             The scheduler to wrap.
         optimizers (one or a list of `torch.optim.Optimizer`):
     """
     def __init__(self, scheduler, optimizers):
         super().__init__(scheduler, optimizers)
         **kwargs:
             Other arguments.
     """
     def __init__(self, params, lr=0.001, weight_decay=0, **kwargs):
         self.params = params
         self.lr = lr
         **kwargs:
             Other arguments.
     """
     def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, lr_scheduler_callable=None, **kwargs):
         self.optimizer = optimizer
         self.total_num_steps = total_num_steps

src/utils/launch.py CHANGED Viewed

@@ -502,7 +502,6 @@ class PrepareForLaunch:
         debug (`bool`, *optional*, defaults to `False`):
             Whether or not this is a debug launch.
     """
     def __init__(self, launcher, distributed_type="NO", debug=False):
         self.launcher = launcher
         self.distributed_type = DistributedType(distributed_type)

         debug (`bool`, *optional*, defaults to `False`):
             Whether or not this is a debug launch.
     """
     def __init__(self, launcher, distributed_type="NO", debug=False):
         self.launcher = launcher
         self.distributed_type = DistributedType(distributed_type)

src/utils/megatron_lm.py CHANGED Viewed

@@ -68,7 +68,6 @@ class MegatronLMDummyDataLoader:
     Args:
         **dataset_kwargs: Megatron data arguments.
     """
     def __init__(self, **dataset_kwargs):
         parser = argparse.ArgumentParser()
         parser = _add_data_args(parser)
@@ -346,7 +345,6 @@ class MegatronLMDummyScheduler:
         **kwargs:
             Other arguments.
     """
     def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, **kwargs):
         self.optimizer = optimizer
         self.total_num_steps = total_num_steps
@@ -392,7 +390,6 @@ class BertTrainStep(AbstractTrainStep):
     Args:
         args (`argparse.Namespace`): Megatron-LM arguments.
     """
     def __init__(self, args):
         super().__init__("BertTrainStep")
         self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
@@ -521,7 +518,6 @@ class GPTTrainStep(AbstractTrainStep):
     Args:
         args (`argparse.Namespace`): Megatron-LM arguments.
     """
     def __init__(self, args):
         super().__init__("GPTTrainStep")
         self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
@@ -627,7 +623,6 @@ class T5TrainStep(AbstractTrainStep):
     Args:
         args (`argparse.Namespace`): Megatron-LM arguments.
     """
     def __init__(self, args):
         super().__init__("T5TrainStep")
         self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
@@ -846,7 +841,6 @@ class MegatronEngine(torch.nn.Module):
         optimizer: Megatron-LM optimizer
         lr_scheduler: Megatron-LM lr scheduler
     """
     def __init__(self, accelerator, model, optimizer, scheduler):
         super(MegatronEngine, self).__init__()
         self.module = model
@@ -892,7 +886,6 @@ class MegatronEngine(torch.nn.Module):
         Args:
             batch_data (:obj:`dict`): The batch data to train on.
         """
         args = get_args()
         timers = get_timers()
@@ -993,7 +986,6 @@ class MegatronEngine(torch.nn.Module):
         Args:
             batch_data (:obj:`dict`): The batch data to evaluate on.
         """
         args = get_args()
         data_chunks = []
         if args.num_micro_batches > 1:
@@ -1176,7 +1168,6 @@ class MegatronEngine(torch.nn.Module):
             length_penalty (float, optional): length penalty for beam search. Defaults to None.
             kwargs: additional key-value arguments
         """
         # checking if required arguments are passed
         args = get_args()
         if args.model_type_name != "gpt":
@@ -1332,7 +1323,6 @@ def avg_losses_across_data_parallel_group(losses):
     Args:
         losses (List[Tensor]): List of losses to average across data parallel group.
     """
     return average_losses_across_data_parallel_group(losses)
@@ -1345,7 +1335,6 @@ def gather_across_data_parallel_groups(tensor):
             The data to gather across data parallel ranks.
     """
     def _gpu_gather_one(tensor):
         if tensor.ndim == 0:
             tensor = tensor.clone()[None]

     Args:
         **dataset_kwargs: Megatron data arguments.
     """
     def __init__(self, **dataset_kwargs):
         parser = argparse.ArgumentParser()
         parser = _add_data_args(parser)
         **kwargs:
             Other arguments.
     """
     def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, **kwargs):
         self.optimizer = optimizer
         self.total_num_steps = total_num_steps
     Args:
         args (`argparse.Namespace`): Megatron-LM arguments.
     """
     def __init__(self, args):
         super().__init__("BertTrainStep")
         self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
     Args:
         args (`argparse.Namespace`): Megatron-LM arguments.
     """
     def __init__(self, args):
         super().__init__("GPTTrainStep")
         self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
     Args:
         args (`argparse.Namespace`): Megatron-LM arguments.
     """
     def __init__(self, args):
         super().__init__("T5TrainStep")
         self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
         optimizer: Megatron-LM optimizer
         lr_scheduler: Megatron-LM lr scheduler
     """
     def __init__(self, accelerator, model, optimizer, scheduler):
         super(MegatronEngine, self).__init__()
         self.module = model
         Args:
             batch_data (:obj:`dict`): The batch data to train on.
         """
         args = get_args()
         timers = get_timers()
         Args:
             batch_data (:obj:`dict`): The batch data to evaluate on.
         """
         args = get_args()
         data_chunks = []
         if args.num_micro_batches > 1:
             length_penalty (float, optional): length penalty for beam search. Defaults to None.
             kwargs: additional key-value arguments
         """
         # checking if required arguments are passed
         args = get_args()
         if args.model_type_name != "gpt":
     Args:
         losses (List[Tensor]): List of losses to average across data parallel group.
     """
     return average_losses_across_data_parallel_group(losses)
             The data to gather across data parallel ranks.
     """
     def _gpu_gather_one(tensor):
         if tensor.ndim == 0:
             tensor = tensor.clone()[None]

src/utils/modeling.py CHANGED Viewed

@@ -395,7 +395,6 @@ def get_non_persistent_buffers(module: nn.Module, recurse: bool = False):
         recurse (`bool`, *optional*, defaults to `False`):
             Whether or not to go look in every submodule or just return the direct non persistent buffers.
     """
     non_persistent_buffers_set = module._non_persistent_buffers_set
     if recurse:
         for _, m in module.named_modules():
@@ -409,7 +408,6 @@ class FindTiedParametersResult(list):
     This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
     a list or on the `values` method as in the future this will be removed.
     """
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -428,7 +426,6 @@ def check_tied_parameters_in_config(model: nn.Module):
     Returns:
         bool: True if the model needs to have tied weights
     """
     # based on model.tie_weights() method
     has_tied_word_embedding = False
     has_tied_encoder_decoder = False

         recurse (`bool`, *optional*, defaults to `False`):
             Whether or not to go look in every submodule or just return the direct non persistent buffers.
     """
     non_persistent_buffers_set = module._non_persistent_buffers_set
     if recurse:
         for _, m in module.named_modules():
     This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
     a list or on the `values` method as in the future this will be removed.
     """
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
     Returns:
         bool: True if the model needs to have tied weights
     """
     # based on model.tie_weights() method
     has_tied_word_embedding = False
     has_tied_encoder_decoder = False

src/utils/offload.py CHANGED Viewed

@@ -85,7 +85,6 @@ class PrefixedDataset(Mapping):
         dataset (`Mapping`): Any map with string keys.
         prefix (`str`): A prefix to add when trying to access any element in the underlying dataset.
     """
     def __init__(self, dataset: Mapping, prefix: str):
         self.dataset = dataset
         self.prefix = prefix
@@ -113,7 +112,6 @@ class OffloadedWeightsLoader(Mapping):
             A dictionary from weight name to their information (`dtype`/ `shape` or safetensors filename). Will default
             to the index saved in `save_folder`.
     """
     def __init__(
         self,
         state_dict: Dict[str, torch.Tensor] = None,

         dataset (`Mapping`): Any map with string keys.
         prefix (`str`): A prefix to add when trying to access any element in the underlying dataset.
     """
     def __init__(self, dataset: Mapping, prefix: str):
         self.dataset = dataset
         self.prefix = prefix
             A dictionary from weight name to their information (`dtype`/ `shape` or safetensors filename). Will default
             to the index saved in `save_folder`.
     """
     def __init__(
         self,
         state_dict: Dict[str, torch.Tensor] = None,

src/utils/operations.py CHANGED Viewed

@@ -150,7 +150,6 @@ def get_data_structure(data):
     Returns:
         The same data structure as `data` with [`~utils.TensorInformation`] instead of tensors.
     """
     def _get_data_structure(tensor):
         return TensorInformation(shape=tensor.shape, dtype=tensor.dtype)
@@ -168,7 +167,6 @@ def get_shape(data):
     Returns:
         The same data structure as `data` with lists of tensor shapes instead of tensors.
     """
     def _get_shape(tensor):
         return list(tensor.shape)
@@ -182,7 +180,6 @@ def initialize_tensors(data_structure):
     Returns:
         The same data structure as `data` with tensors instead of [`~utils.TensorInformation`].
     """
     def _initialize_tensor(tensor_info):
         return torch.empty(*tensor_info.shape, dtype=tensor_info.dtype)
@@ -222,7 +219,6 @@ def listify(data):
     Returns:
         The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
     """
     def _convert_to_list(tensor):
         tensor = tensor.detach().cpu()
         if tensor.dtype == torch.bfloat16:
@@ -293,7 +289,6 @@ class DistributedOperationException(Exception):
     An exception class for distributed operations. Raised if the operation cannot be performed due to the shape of the
     tensors.
     """
     pass
@@ -301,7 +296,6 @@ def verify_operation(function):
     """
     Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
     """
     @wraps(function)
     def wrapper(*args, **kwargs):
         if PartialState().distributed_type == DistributedType.NO or not PartialState().debug:
@@ -337,7 +331,6 @@ def chained_operation(function):
     Checks that `verify_operation` failed and if so reports a more helpful error chaining the existing
     `DistributedOperationException`.
     """
     @wraps(function)
     def wrapper(*args, **kwargs):
         try:
@@ -469,7 +462,6 @@ def slice_tensors(data, tensor_slice, process_index=None, num_processes=None):
     Returns:
         The same data structure as `data` with all the tensors slices.
     """
     def _slice_tensor(tensor, tensor_slice):
         return tensor[tensor_slice]
@@ -518,7 +510,6 @@ def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
         pad_first (`bool`, *optional*, defaults to `False`):
             Whether to pad at the beginning or the end.
     """
     def _pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
         if getattr(tensor, "is_nested", False):
             warnings.warn(
@@ -572,7 +563,6 @@ def reduce(tensor, reduction="mean", scale=1.0):
     Returns:
         The same data structure as `data` with all the tensors reduced.
     """
     def _reduce_across_processes(tensor, reduction="mean", scale=1.0):
         state = PartialState()
         cloned_tensor = tensor.clone()
@@ -602,7 +592,6 @@ def convert_to_fp32(tensor):
     Returns:
         The same data structure as `tensor` with all tensors that were in FP16/BF16 precision converted to FP32.
     """
     def _convert_to_fp32(tensor):
         return tensor.float()
@@ -624,7 +613,6 @@ class ConvertOutputsToFp32:
     Returns:
         The same function as `model_forward` but with converted outputs.
     """
     def __init__(self, model_forward):
         self.model_forward = model_forward
         update_wrapper(self, model_forward)

     Returns:
         The same data structure as `data` with [`~utils.TensorInformation`] instead of tensors.
     """
     def _get_data_structure(tensor):
         return TensorInformation(shape=tensor.shape, dtype=tensor.dtype)
     Returns:
         The same data structure as `data` with lists of tensor shapes instead of tensors.
     """
     def _get_shape(tensor):
         return list(tensor.shape)
     Returns:
         The same data structure as `data` with tensors instead of [`~utils.TensorInformation`].
     """
     def _initialize_tensor(tensor_info):
         return torch.empty(*tensor_info.shape, dtype=tensor_info.dtype)
     Returns:
         The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
     """
     def _convert_to_list(tensor):
         tensor = tensor.detach().cpu()
         if tensor.dtype == torch.bfloat16:
     An exception class for distributed operations. Raised if the operation cannot be performed due to the shape of the
     tensors.
     """
     pass
     """
     Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
     """
     @wraps(function)
     def wrapper(*args, **kwargs):
         if PartialState().distributed_type == DistributedType.NO or not PartialState().debug:
     Checks that `verify_operation` failed and if so reports a more helpful error chaining the existing
     `DistributedOperationException`.
     """
     @wraps(function)
     def wrapper(*args, **kwargs):
         try:
     Returns:
         The same data structure as `data` with all the tensors slices.
     """
     def _slice_tensor(tensor, tensor_slice):
         return tensor[tensor_slice]
         pad_first (`bool`, *optional*, defaults to `False`):
             Whether to pad at the beginning or the end.
     """
     def _pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
         if getattr(tensor, "is_nested", False):
             warnings.warn(
     Returns:
         The same data structure as `data` with all the tensors reduced.
     """
     def _reduce_across_processes(tensor, reduction="mean", scale=1.0):
         state = PartialState()
         cloned_tensor = tensor.clone()
     Returns:
         The same data structure as `tensor` with all tensors that were in FP16/BF16 precision converted to FP32.
     """
     def _convert_to_fp32(tensor):
         return tensor.float()
     Returns:
         The same function as `model_forward` but with converted outputs.
     """
     def __init__(self, model_forward):
         self.model_forward = model_forward
         update_wrapper(self, model_forward)