# Copyright 2021 MosaicML. All Rights Reserved.
"""Train models!
The trainer supports models with :class:`~composer.models.base.ComposerModel` instances.
The :class:`.Trainer` is highly customizable and can
support a wide variety of workloads.
Example
--------
Train a model and save a checkpoint:
.. testcode::
import os
from composer import Trainer
### Create a trainer
trainer = Trainer(
model=model,
train_dataloader=train_dataloader,
max_duration="1ep",
eval_dataloader=eval_dataloader,
optimizers=optimizer,
schedulers=scheduler,
device="cpu",
validate_every_n_epochs=1,
save_folder="checkpoints",
save_filename="ep{epoch}.pt",
save_interval="1ep",
save_overwrite=True,
)
# Fit and run evaluation for 1 epoch.
# Save a checkpoint after 1 epoch as specified during trainer creation.
trainer.fit()
Load the checkpoint and resume training:
.. testcode::
# Get the saved checkpoint filepath
checkpoint_path = trainer.saved_checkpoints.pop()[0]
# Create a new trainer with the `load_path` argument set to the checkpoint path.
trainer = Trainer(
model=model,
train_dataloader=train_dataloader,
max_duration="2ep",
eval_dataloader=eval_dataloader,
optimizers=optimizer,
schedulers=scheduler,
device="cpu",
validate_every_n_epochs=1,
load_path=checkpoint_path,
)
# Continue training and running evaluation where the previous trainer left off
# until the new max_duration is reached.
# In this case it will be one additional epoch to reach 2 epochs total.
trainer.fit()
"""
from __future__ import annotations
import contextlib
import datetime
import itertools
import logging
import pathlib
import sys
import textwrap
import warnings
from typing import Any, Callable, ContextManager, Dict, List, Optional, Sequence, TextIO, Tuple, Union, cast
import torch
import torch.distributed
import torch.utils.data
from torch.cuda.amp.grad_scaler import GradScaler
from torch.nn.parallel import DistributedDataParallel
from torchmetrics import Metric, MetricCollection
import composer
from composer.algorithms import ScaleSchedule
from composer.callbacks import CheckpointSaver
from composer.core import Algorithm, Callback, DataSpec, Engine, Evaluator, Event, Precision, State, Time, Timestamp
from composer.core.types import Batch, BreakEpochException, DataLoader, PyTorchScheduler
from composer.datasets.dataloader import unwrap_data_loader
from composer.loggers import Logger, LoggerDestination, LogLevel, ProgressBarLogger
from composer.models.base import ComposerModel
from composer.optim.decoupled_weight_decay import DecoupledSGDW
from composer.optim.scheduler import ComposerScheduler, compile_composer_scheduler
from composer.profiler import DataLoaderProfiler, Profiler, ProfilerAction, SystemProfiler, TorchProfiler, TraceHandler
from composer.trainer._deepspeed import _fix_batch_precision_for_deepspeed, _parse_deepspeed_config
from composer.trainer._scale_schedule import scale_pytorch_scheduler
from composer.trainer._scaler import ClosureGradScaler
from composer.trainer.ddp import DDPSyncStrategy, _ddp_sync_context, _prepare_ddp_module
from composer.trainer.devices import Device, DeviceCPU, DeviceGPU
from composer.utils import dist, ensure_tuple, map_collection, module_surgery, reproducibility
from composer.utils.checkpoint import load_checkpoint, save_checkpoint
from composer.utils.import_helpers import MissingConditionalImportError
from composer.utils.object_store import ObjectStore
log = logging.getLogger(__name__)
__all__ = ["Trainer"]
[docs]class Trainer:
"""Trainer for training a models with Composer algorithms. See the Trainer guide for more information.
Args:
model (ComposerModel): The model to train. Can be user-defined or one of the models included
with Composer.
.. seealso:: :mod:`composer.models` for models built into Composer.
train_dataloader (DataLoader, DataSpec, or dict): The :class:`.DataLoader`, :class:`.DataSpec`,
or dict of :class:`.DataSpec` kwargs for the training data. In order to specify custom
preprocessing steps on each data batch, specify a :class:`.DataSpec` instead of a
:class:`.DataLoader`.
.. note:: The ``train_dataloader`` should yield per-rank batches. Each per-rank batch
will then be further divided based on the ``grad_accum`` parameter. For example, if the
desired optimization batch size is ``2048`` and training is happening across 8 GPUs, then each
``train_dataloader`` should yield a batch of size ``2048 / 8 = 256``. If ``grad_accum = 2``,
then the per-rank batch will be divided into microbatches of size ``256 / 2 = 128``.
max_duration (int, str, or Time): The maximum duration to train. Can be an integer, which will be
interpreted to be epochs, a str (e.g. ``1ep``, or ``10ba``), or a :class:`.Time` object.
eval_dataloader (DataLoader | DataSpec | Evaluator | Sequence[Evaluator], optional): The :class:`.DataLoader`,
:class:`.DataSpec`, :class:`.Evaluator`, or sequence of evaluators for the evaluation data.
To evaluate one or more specific metrics across one or more datasets, pass in an
:class:`.Evaluator`. If a :class:`.DataSpec` or :class:`.DataLoader` is passed in, then all
metrics returned by ``model.metrics()`` will be used during evaluation.
``None`` results in no evaluation. (default: ``None``)
algorithms (Algorithm | Sequence[Algorithm], optional): The algorithms to use during training. If ``None``, then
no algorithms will be used. (default: ``None``)
.. seealso:: :mod:`composer.algorithms` for the different algorithms built into Composer.
optimizers (torch.optim.Optimizer, optional): The optimizer.
If ``None``, will be set to ``DecoupledSGDW(model.parameters(), lr=0.1)``. (default: ``None``)
.. seealso:: :mod:`composer.optim` for the different optimizers built into Composer.
schedulers (PyTorchScheduler | ComposerScheduler | Sequence[PyTorchScheduler | ComposerScheduler], optional):
The learning rate schedulers. If ``[]`` or ``None``, the learning rate will be constant.
(default: ``None``).
.. seealso:: :mod:`composer.optim.scheduler` for the different schedulers built into Composer.
device (str or Device, optional): The device to use for training. Either ``cpu`` or ``gpu``.
(default: ``cpu``)
grad_accum (Union[int, str], optional): The number of microbatches to split a per-device batch into. Gradients
are summed over the microbatches per device. If set to ``auto``, dynamically increases grad_accum
if microbatch is too large for GPU. (default: ``1``)
.. note:: This is implemented by taking the batch yielded by the ``train_dataloader`` and splitting
it into ``grad_accum`` sections. Each section is of size ``train_dataloader // grad_accum``.
If the batch size of the dataloader is not divisible by ``grad_accum``,
then the last section will be of size ``batch_size % grad_accum``.
grad_clip_norm (float, optional): The norm to clip gradient magnitudes to. Set to ``None`` for no gradient
clipping. (default: ``None``)
validate_every_n_batches (int, optional): Compute metrics on evaluation data every N batches.
Set to ``-1`` to never validate on a batchwise frequency. (default: ``-1``)
validate_every_n_epochs (int, optional): Compute metrics on evaluation data every N epochs.
Set to ``-1`` to never validate on a epochwise frequency. (default: ``1``)
compute_training_metrics (bool, optional): ``True`` to compute metrics on training data and ``False`` to not.
(default: ``False``)
precision (str or Precision, optional): Numerical precision to use for training. One of ``fp32``, ``fp16``
or ``amp`` (recommended). (default: ``Precision.FP32``)
.. note::
``fp16`` only works if ``deepspeed_config`` is also provided.
scale_schedule_ratio (float, optional): Ratio by which to scale the training duration and learning rate
schedules. E.g., ``0.5`` makes the schedule take half as many epochs and ``2.0`` makes it take twice as
many epochs. ``1.0`` means no change. (default: ``1.0``)
.. note ::
Training for less time, while rescaling the learning rate schedule,
is a strong baseline approach to speeding up training. E.g., training
for half duration often yields minor accuracy degradation,
provided that the learning rate schedule is also rescaled to take half as long.
To see the difference, consider training for half as long using a cosine
annealing learning rate schedule. If the schedule is not rescaled,
training ends while the learning rate is still ~0.5 of the initial LR.
If the schedule is rescaled with ``scale_schedule_ratio``, the LR schedule
would finish the entire cosine curve, ending with a learning rate near zero.
step_schedulers_every_batch (bool, optional): By default, native
`PyTorch schedulers <https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate>`_
are updated every epoch, while :doc:`Composer Schedulers</trainer/schedulers>` are updated every step.
Setting this to ``True`` will force schedulers to be stepped every batch,
while ``False`` means schedulers stepped every epoch. ``None`` indicates the default behavior.
(default: ``None``)
dist_timeout (float, optional): Timeout, in seconds, for initializing the distributed process group.
(default: ``15.0``)
ddp_sync_strategy (str or DDPSyncStrategy, optional): The strategy to use for synchronizing gradients.
Leave unset to let the trainer auto-configure this. See :class:`.DDPSyncStrategy`
for more details.
seed (int, optional): The seed used in randomization. If ``None``, then a random seed
will be created. (default: ``None``)
.. note:: In order to get reproducible results, call the
:func:`.seed_all` function at the start of your script with the seed
passed to the trainer. This will ensure any initialization done before the trainer init
(ex. model weight initialization) also uses the provided seed.
.. seealso:: :mod:`composer.utils.reproducibility` for more details on reproducibility.
deterministic_mode (bool, optional): Run the model deterministically. (default: ``False``)
.. note:: This is an experimental feature. Performance degradations expected. Certain Torch modules may
not have deterministic implementations, which will result in a crash.
.. note:: In order to get reproducible results, call the
:func:`.configure_deterministic_mode` function at the start of your script.
This will ensure any initialization done before the trainer init also runs deterministically.
.. seealso:: :mod:`composer.utils.reproducibility` for more details on reproducibility.
run_name (str, optional): A name for this training run. If not specified, one will be generated automatically.
.. seealso:: :class:`~composer.loggers.logger.Logger`
loggers (LoggerDestination | Sequence[LoggerDestination], optional): The destinations to log training information to.
If ``None``, will be set to ``[ProgressBarLogger()]``. (default: ``None``)
.. seealso:: :mod:`composer.loggers` for the different loggers built into Composer.
progress_bar (bool, optional): Whether to show a progress bar. (default: ``True``)
log_to_console (bool, optional): Whether to print logging statements to the console. (default: ``None``)
The default behavior (when set to ``None``) only prints logging statements when ``show_pbar`` is ``False``.
console_log_level (LogLevel | str | (State, LogLevel) -> bool, optional): The maximum log level which
should be printed to the console. (default: :attr:`.LogLevel.EPOCH`)
It can either be :class:`.LogLevel`, a string corresponding to a :class:`.LogLevel`, or a callable
that takes the training :class:`.State` and the :class:`.LogLevel` and returns a boolean of whether this
statement should be printed.
This parameter has no effect if ``log_to_console`` is ``False``, or is unspecified and ``progres_bar`` is
``True``.
console_stream (TextIO | str, optional): The stream to write to. If a string, it can either be
``'stdout'`` or ``'stderr'``. (default: :attr:`sys.stderr`)
callbacks (Callback | Sequence[Callback], optional): The callbacks to run during training. If ``None``,
then no callbacks will be run. (default: ``None``).
.. seealso:: :mod:`composer.callbacks` for the different callbacks built into Composer.
load_path (str, optional): The path format string to an existing checkpoint file.
It can be a path to a file on the local disk, a URL, or if ``load_object_store`` is set, the object name
for a checkpoint in a cloud bucket.
When using `Deepspeed ZeRO <https://www.deepspeed.ai/tutorials/zero/>`_, checkpoints are shareded by rank.
Instead of hard-coding the rank in the ``path``, use the following format variables:
+------------------------+-------------------------------------------------------+
| Variable | Description |
+========================+=======================================================+
| ``{rank}`` | The global rank, as returned by |
| | :func:`~.dist.get_global_rank`. |
+------------------------+-------------------------------------------------------+
| ``{local_rank}`` | The local rank of the process, as returned by |
| | :func:`~.dist.get_local_rank`. |
+------------------------+-------------------------------------------------------+
| ``{node_rank}`` | The node rank, as returned by |
| | :func:`~.dist.get_node_rank`. |
+------------------------+-------------------------------------------------------+
For example, suppose that checkpoints are stored in the following structure:
.. code-block::
my_model/ep1-rank0.tar
my_model/ep1-rank1.tar
my_model/ep1-rank2.tar
...
Then, ``load_path`` should be set to ``my_model/ep1-rank{rank}.tar``, and all ranks will load the
correct state.
If ``None`` then no checkpoint will be loaded. (default: ``None``)
load_object_store (ObjectStore, optional): If the ``load_path`` is in an object store
(i.e. AWS S3 or Google Cloud Storage), an instance of :class:`.ObjectStore` which
will be used to retreive the checkpoint. Otherwise, if the checkpoint is a local filepath,
set to ``None``. Ignored if ``load_path`` is ``None``. (default: ``None``)
Example:
.. testsetup::
import composer.trainer
composer.trainer.trainer.load_checkpoint = lambda *args, **kwargs: None
.. testcode::
from composer import Trainer
from composer.utils import ObjectStore
# Create the object store provider with the specified credentials
creds = {"key": "object_store_key",
"secret": "object_store_secret"}
store = ObjectStore(provider="s3",
container="my_container",
provider_kwargs=creds)
checkpoint_path = "./path_to_the_checkpoint_in_object_store"
# Create a trainer which will load a checkpoint from the specified object store
trainer = Trainer(
model=model,
train_dataloader=train_dataloader,
max_duration="10ep",
eval_dataloader=eval_dataloader,
optimizers=optimizer,
schedulers=scheduler,
device="cpu",
validate_every_n_epochs=1,
load_path=checkpoint_path,
load_object_store=store,
)
.. testcleanup::
trainer.engine.close()
load_weights_only (bool, optional): Whether or not to only restore the weights from the checkpoint without
restoring the associated state. Ignored if ``load_path`` is ``None``. (default: ``False``)
load_strict (bool, optional): Ensure that the set of weights in the checkpoint and model must exactly match.
Ignored if ``load_path`` is ``None``. (default: ``False``)
load_chunk_size (int, optional): Chunk size (in bytes) to use when downloading checkpoints.
Ignored if ``load_path`` is either ``None`` or a local file path. (default: ``1,048,675``)
load_progress_bar (bool, optional): Display the progress bar for downloading the checkpoint.
Ignored if ``load_path`` is either ``None`` or a local file path. (default: ``True``)
save_folder (str, optional): Format string for the folder where checkpoints are saved.
If ``None``, checkpoints will not be saved. (default: ``None``)
.. seealso:: :class:`~.CheckpointSaver`
.. note::
For fine-grained control on checkpoint saving (e.g. to save different types of checkpoints
at different intervals), leave this parameter as ``None``, and instead pass
instance(s) of :class:`~.CheckpointSaver` directly as ``callbacks``.
save_filename (str, optional): A format string describing how to name checkpoints.
This parameter has no effect if ``save_folder`` is ``None``.
(default: ``"ep{epoch}-ba{batch}-rank{rank}"``)
.. seealso:: :class:`~.CheckpointSaver`
save_artifact_name (str, optional): A format string describing how to name checkpoints in loggers.
This parameter has no effect if ``save_folder`` is ``None``.
(default: ``"{run_name}/checkpoints/ep{epoch}-ba{batch}-rank{rank}"``)
.. seealso:: :class:`~.CheckpointSaver`
save_latest_filename (str, optional): A format string for the name of a symlink
(relative to ``save_folder``) that points to the last saved checkpoint.
This parameter has no effect if ``save_folder`` is ``None``.
To disable symlinking, set this to ``None``. (default: ``"latest-rank{rank}"``)
.. seealso:: :class:`~.CheckpointSaver`
save_latest_artifact_name (str, optional): A format string describing how to name symlinks in loggers.
This parameter has no effect if ``save_folder``, ``save_latest_filename``, or ``save_artifact_name`` are ``None``.
To disable symlinking in logger, set this or ``save_latest_filename`` to ``None``. (default: ``"{run_name}/checkpoints/latest-rank{rank}"``)
.. seealso:: :class:`~.CheckpointSaver`
save_overwrite (bool, optional): Whether existing checkpoints should be overridden.
This parameter has no effect if ``save_folder`` is None. (default: ``False``)
.. seealso:: :class:`~.CheckpointSaver`
save_interval (Time | str | int | (State, Event) -> bool): A :class:`Time`, time-string, integer (in epochs),
or a function that takes (state, event) and returns a boolean whether a checkpoint should be saved.
This parameter has no effect if ``save_folder`` is ``None``. (default: ``'1ep'``)
.. seealso:: :class:`~.CheckpointSaver`
save_weights_only (bool, optional): Whether to save only the model weights instead of the entire training
state. This parameter has no effect if ``save_folder`` is ``None``. (default: ``False``)
.. seealso:: :class:`~.CheckpointSaver`
save_num_checkpoints_to_keep (int, optional): The number of checkpoints to keep locally. The oldest checkpoints
are removed first. Set to ``-1`` to keep all checkpoints locally. (default: ``-1``)
Checkpoints will be removed after they have been logged as a file artifact. For example, when this callback
is used in conjunction with the :class:`~composer.loggers.object_store_logger.ObjectStoreLogger`, set this
parameter to ``0`` to immediately delete checkpoints from the local disk after they have been uploaded to
the object store.
This parameter only controls how many checkpoints are kept locally; checkpoints are not deleted from
artifact stores.
train_subset_num_batches (int, optional): If specified, finish every epoch early after training
on this many batches. This parameter has no effect if it is greater than ``len(train_dataloader)``.
If ``None``, then the entire dataloader will be iterated over. (default: ``None``)
eval_subset_num_batches (int, optional): If specified, evaluate on this many batches.
This parameter has no effect if it is greater than ``len(eval_dataloader)``.
If ``None``, then the entire dataloader will be iterated over. (default: ``None``)
deepspeed_config (bool or Dict[str, Any], optional): Configuration for DeepSpeed, formatted as a JSON
according to `DeepSpeed's documentation <https://www.deepspeed.ai/docs/config-json/>`_. If ``True`` is
provided, the trainer will initialize the DeepSpeed engine with an empty config ``{}``. If ``False``
is provided, deepspeed will not be used. (default: ``False``)
prof_schedule ((State) -> ProfilerAction, optional): The profiler scheduler.
Must be specified in conjunction with ``prof_trace_handlers`` to use the profiler.
.. testcode::
from composer.trainer import Trainer
from composer.profiler import JSONTraceHandler, cyclic_schedule
trainer = Trainer(
...,
prof_trace_handlers=JSONTraceHandler(
folder='traces',
),
prof_schedule=cyclic_schedule(
skip_first=1,
wait=0,
warmup=1,
active=4,
repeat=1,
),
)
.. testcleanup::
trainer.engine.close()
.. seealso:: :mod:`composer.profiler` for more details on profiling with the trainer.
prof_trace_handlers (TraceHandler | Sequence[TraceHandler], optional): Profiler trace handlers.
Must be specified in conjunction with ``prof_trace_handlers`` to use the profiler.
.. seealso:: :mod:`composer.profiler` for more details on profiling with the trainer.
sys_prof_cpu (bool, optional): Whether to record cpu statistics.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified. (default: ``True``).
sys_prof_memory (bool, optional): Whether to record memory statistics.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified. (default: ``False``).
sys_prof_disk (bool, optional): Whether to record disk statistics.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified. (default: ``False``).
sys_prof_net (bool, optional): Whether to record network statistics.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified. (default: ``False``).
sys_prof_stats_thread_interval_seconds (float, optional): Interval to record stats, in seconds.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified. (default: ``0.5``).
torch_prof_folder (str, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_filename (str, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_artifact_name (str, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_overwrite (bool, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_use_gzip (bool, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_record_shapes (bool, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_profile_memory (bool, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_with_stack (bool, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_with_flops (bool, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
torch_prof_num_traces_to_keep (int, optional): See :class:`~composer.profiler.torch_profiler.TorchProfiler`.
Ignored if ``prof_schedule`` and ``prof_trace_handlers`` are not specified.
Attributes:
state (State): The :class:`.State` object used to store training state.
evaluators (List[Evaluator]): The :class:`.Evaluator` objects to use for validation
during training.
logger (Logger): The :class:`.Logger` used for logging.
engine (Engine): The :class:`.Engine` used for running callbacks and algorithms.
"""
def __init__(
self,
*,
model: ComposerModel,
train_dataloader: Union[DataLoader, DataSpec],
max_duration: Union[int, str, Time],
eval_dataloader: Optional[Union[DataLoader, DataSpec, Evaluator, Sequence[Evaluator]]] = None,
algorithms: Optional[Union[Algorithm, Sequence[Algorithm]]] = None,
optimizers: Optional[torch.optim.Optimizer] = None,
schedulers: Optional[Union[ComposerScheduler, Sequence[ComposerScheduler]]] = None,
# device
device: Optional[Union[str, Device]] = None,
# training hparams
grad_accum: Union[int, str] = 1,
grad_clip_norm: Optional[float] = None,
validate_every_n_batches: int = -1,
validate_every_n_epochs: int = 1,
compute_training_metrics: bool = False,
precision: Union[str, Precision] = Precision.FP32,
scale_schedule_ratio: float = 1.0,
step_schedulers_every_batch: Optional[bool] = None,
# dist hparams
dist_timeout: float = 300.0,
ddp_sync_strategy: Optional[Union[str, DDPSyncStrategy]] = None,
# randomness
seed: Optional[int] = None,
deterministic_mode: bool = False,
# logging and callbacks
run_name: Optional[str] = None,
loggers: Optional[Union[LoggerDestination, Sequence[LoggerDestination]]] = None,
callbacks: Union[Callback, Sequence[Callback]] = tuple(),
progress_bar: bool = True,
log_to_console: Optional[bool] = None,
console_log_level: Union[LogLevel, str, Callable[[State, LogLevel], bool]] = LogLevel.EPOCH,
console_stream: Union[str, TextIO] = sys.stderr,
# load checkpoint
load_path: Optional[str] = None,
load_object_store: Optional[ObjectStore] = None,
load_weights_only: bool = False,
load_strict: bool = False,
load_chunk_size: int = 1_048_576,
load_progress_bar: bool = True,
# save_checkpoint
save_folder: Optional[str] = None,
save_filename: str = "ep{epoch}-ba{batch}-rank{rank}",
save_artifact_name: str = "{run_name}/checkpoints/ep{epoch}-ba{batch}-rank{rank}",
save_latest_filename: str = "latest-rank{rank}",
save_latest_artifact_name: str = "{run_name}/checkpoints/latest-rank{rank}",
save_overwrite: bool = False,
save_interval: Union[str, int, Time, Callable[[State, Event], bool]] = "1ep",
save_weights_only: bool = False,
save_num_checkpoints_to_keep: int = -1,
# subset parameters
train_subset_num_batches: Optional[int] = None,
eval_subset_num_batches: Optional[int] = None,
# DeepSpeed
deepspeed_config: Union[bool, Dict[str, Any]] = False,
# profiling
prof_trace_handlers: Optional[Union[TraceHandler, Sequence[TraceHandler]]] = None,
prof_schedule: Optional[Callable[[State], ProfilerAction]] = None,
sys_prof_cpu: bool = True,
sys_prof_memory: bool = False,
sys_prof_disk: bool = False,
sys_prof_net: bool = False,
sys_prof_stats_thread_interval_seconds: float = 0.5,
torch_prof_folder: str = '{run_name}/torch_traces',
torch_prof_filename: str = 'rank{rank}.{batch}.pt.trace.json',
torch_prof_artifact_name: str = '{run_name}/torch_traces/rank{rank}.{batch}.pt.trace.json',
torch_prof_overwrite: bool = False,
torch_prof_use_gzip: bool = False,
torch_prof_record_shapes: bool = False,
torch_prof_profile_memory: bool = True,
torch_prof_with_stack: bool = False,
torch_prof_with_flops: bool = True,
torch_prof_num_traces_to_keep: int = -1,
):
# surpressing GradScaler warnings as they are always created
# self._use_grad_scaling() will raise a RuntimeError if grad scaling is not available when it is required
warnings.filterwarnings(action="ignore", message="torch.cuda.amp.GradScaler")
# ScaleSchedule is a deprecated algorithm, but if it is used, updated SSR with its ratio.
# TODO(#434): Remove this completely.
for algorithm in ensure_tuple(algorithms):
if isinstance(algorithm, ScaleSchedule):
scale_schedule_ratio = algorithm.ratio
if isinstance(max_duration, str):
max_duration = Time.from_timestring(max_duration)
elif isinstance(max_duration, int):
max_duration = Time.from_epoch(max_duration)
orig_max_duration = max_duration
if scale_schedule_ratio != 1.0:
max_duration = cast(Time[int], orig_max_duration * scale_schedule_ratio)
log.info(f'max_duration changed from {orig_max_duration} to {max_duration}')
if max_duration.value == 0:
raise ValueError(
'Scale schedule has reduced the max_duration to 0. Set a higher ratio or use more epochs.')
if isinstance(deepspeed_config, bool):
self._deepspeed_config = {} if deepspeed_config else None
else:
self._deepspeed_config = deepspeed_config
if not device:
self._device = DeviceCPU() if not self.deepspeed_enabled else DeviceGPU()
elif isinstance(device, str):
if device == 'cpu':
self._device = DeviceCPU()
elif device == 'gpu':
self._device = DeviceGPU()
else:
raise ValueError(f'device ({device}) must be one of (cpu, gpu).')
else:
if not isinstance(device, Device):
raise ValueError('device must be of class Device')
self._device = device
if self.deepspeed_enabled or dist.get_world_size() > 1:
# deepspeed requires torch.distributed to be initialized, even if the world size is 1
# distributed is always required with multi-rank training
dist.initialize_dist(self._device.dist_backend, datetime.timedelta(seconds=dist_timeout))
if not seed:
seed = reproducibility.get_random_seed()
# Ensure that each process has a seed = rank_zero_seed + global_rank
# This "deterministically different" seed behavior is required to be able
# to restore seeds when resuming form checkpoints, since only the
# `rank_zero_seed` is stored on state.
if seed < 0 or seed > reproducibility.MAX_SEED:
raise ValueError(f"Invalid seed: {seed}. It must be on [0; 2**32 - 1)")
rank_zero_seed = self._device.tensor_to_device(torch.tensor(
[seed], dtype=torch.int64)) # using int64 to prevent overflow
dist.broadcast(rank_zero_seed, src=0)
rank_zero_seed = rank_zero_seed.item()
assert isinstance(rank_zero_seed, int)
seed = rank_zero_seed + dist.get_global_rank()
log.info(f"Setting seed to {seed}")
# If hparams is used to create the Trainer this function is called twice
# which is okay because all runs with the hparams codepath will do this
reproducibility.seed_all(seed)
# some algorithms require specific settings
self._backwards_create_graph = any(map(lambda x: x.backwards_create_graph, ensure_tuple(algorithms)))
find_unused_parameters = any(map(lambda x: x.find_unused_parameters, ensure_tuple(algorithms)))
self._find_unused_parameters = find_unused_parameters
if ddp_sync_strategy is None:
self._ddp_sync_strategy = DDPSyncStrategy.SINGLE_AUTO_SYNC if not find_unused_parameters else DDPSyncStrategy.FORCED_SYNC
else:
self._ddp_sync_strategy = DDPSyncStrategy(ddp_sync_strategy)
# convert eval_dataloader to `List[Evaluator]`
self.evaluators: List[Evaluator] = []
for evaluator in ensure_tuple(eval_dataloader):
if isinstance(evaluator, Evaluator):
self.evaluators.append(evaluator)
else:
metrics = model.metrics(train=False)
evaluator = Evaluator(label="eval", dataloader=evaluator, metrics=metrics)
self.evaluators.append(evaluator)
self._eval_subset_num_batches = eval_subset_num_batches
# do a check here to make sure there is at least one validation set
if len(self.evaluators) == 0:
warnings.warn(
textwrap.dedent("""No evaluation dataset was specified. Please specify `eval_dataloader` to periodically
evaluate your model while training."""),
category=UserWarning)
if not isinstance(train_dataloader, DataSpec):
train_dataloader = DataSpec(train_dataloader)
self._train_data_spec = train_dataloader
unwrapped_data_loader = unwrap_data_loader(self._train_data_spec.dataloader)
if isinstance(unwrapped_data_loader, torch.utils.data.DataLoader):
if unwrapped_data_loader._iterator is not None:
raise ValueError(
textwrap.dedent("""\
The `train_dataloader` has an active iterator. This could occur
if `persistent_workers=True` and the dataloader has already been iterated,
or if the dataloader is mid-epoch. It is required that the training dataloader
does not have an active iterator, so CPU dataset augmentations can be
correctly inserted.
To fix, please do not iterate over the dataloader before passing it into
the trainer."""))
# TODO(#123): DeepSpeed still needs a precision context, but it's not completely clear how to
# handle this with our version of Pytorch
precision_context = self._device.precision_context if not self.deepspeed_enabled else cast(
Callable[..., ContextManager], contextlib.nullcontext)
if isinstance(precision, str):
precision = Precision(precision)
# optimizers and schedulers
if not optimizers:
optimizers = DecoupledSGDW(list(model.parameters()), lr=0.1)
warnings.warn(f"No optimizer was specified. Defaulting to {repr(optimizers)}")
num_optimizers = len(ensure_tuple(optimizers))
if num_optimizers != 1:
raise NotImplementedError(f"Only one optimizer is supported; found {num_optimizers} optimizers")
# Set initial grad_accum to 1 if using adaptive
self.adaptive_grad_accum = grad_accum == "auto"
if self.adaptive_grad_accum:
grad_accum = 1
warnings.warn(textwrap.dedent("""Setting `grad_accum='auto'` is an experimental feature
which may cause uncaught Cuda Out of Memory errors. In this case, please manually
set grad_accum explicitly to an integer instead.
"""),
category=UserWarning)
# Cannot use adaptive grad accum on CPU
if isinstance(self._device, DeviceCPU) and self.adaptive_grad_accum:
raise ValueError("Cannot use adaptive grad_accum on CPU. Please set grad_accum >= 1")
# grad_accum should be int as we've already handled "auto" case
if isinstance(grad_accum, str):
raise ValueError("grad_accum must be an int or ``auto``")
self.state = State(
max_duration=max_duration,
rank_zero_seed=rank_zero_seed,
algorithms=algorithms,
model=model,
callbacks=callbacks,
grad_accum=grad_accum,
precision=precision,
precision_context=precision_context,
train_dataloader=train_dataloader.dataloader,
evaluators=self.evaluators,
optimizers=optimizers,
steps_per_epoch=train_subset_num_batches,
)
pytorch_schedulers = [
scheduler for scheduler in ensure_tuple(schedulers) if isinstance(scheduler, PyTorchScheduler)
]
if len(pytorch_schedulers) > 0:
if step_schedulers_every_batch is True:
log.info(
textwrap.dedent(f"""\
Schedulers are being steped every batch, as `step_schedulers_every_batch` is True.
The trainer cannot automatically convert the parameters (e.g. step_size, T_max) of the
PyTorch {type(pytorch_schedulers[0]).__name__} scheduler to be in terms of batches.
Please ensure that the scheduler parameters are in terms of batches, not epochs.
Alternatively, use a ComposerScheduler. For more information, see
https://docs.mosaicml.com/en/v{composer.__version__}/trainer/schedulers.html.
"""))
if step_schedulers_every_batch is None:
# only if all schedulers are ComposerSchedulers, then we can step every batch by default
step_schedulers_every_batch = False
log.info(
textwrap.dedent(f"""\
Schedulers will be stepped every epoch because the Trainer was constructed with a PyTorch
{type(pytorch_schedulers[0]).__name__} scheduler. To step the schedulers every batch, adjust the
scheduler parameters (e.g. step_size, T_max) to be in terms of batches and set
`step_schedulers_every_batch` to True, or alternatively use a ComposerScheduler. For more information,
see https://docs.mosaicml.com/en/v{composer.__version__}/trainer/schedulers.html."""))
else:
step_schedulers_every_batch = True
self._step_schedulers_every_batch = step_schedulers_every_batch
for scheduler in ensure_tuple(schedulers):
if isinstance(scheduler, PyTorchScheduler):
scale_pytorch_scheduler(scheduler, scale_schedule_ratio)
self.state.schedulers.append(scheduler)
else: # it's a composer scheduler
self.state.schedulers.append(compile_composer_scheduler(scheduler, self.state, scale_schedule_ratio))
if len(self.state.schedulers) == 0:
warnings.warn(f"NoSchedulerWarning: No schedulers were specified. The learning rate will be constant.")
# Configure profilers if profiling is enabled
if prof_schedule is not None or len(ensure_tuple(prof_trace_handlers)) > 0:
if prof_schedule is None or len(ensure_tuple(prof_trace_handlers)) == 0:
raise ValueError(
"To use the profiler, both `prof_schedule` and `prof_trans_handlers` must be specified.")
self.state.profiler = Profiler(state=self.state,
trace_handlers=ensure_tuple(prof_trace_handlers),
schedule=prof_schedule)
self.state.callbacks.append(DataLoaderProfiler())
if sys_prof_cpu or sys_prof_memory or sys_prof_disk or sys_prof_net:
self.state.callbacks.append(
SystemProfiler(profile_cpu=sys_prof_cpu,
profile_memory=sys_prof_memory,
profile_disk=sys_prof_disk,
profile_net=sys_prof_net,
stats_thread_interval_seconds=sys_prof_stats_thread_interval_seconds))
if torch_prof_record_shapes or torch_prof_profile_memory or torch_prof_with_stack or torch_prof_with_flops:
self.state.callbacks.append(
TorchProfiler(filename=torch_prof_filename,
folder=torch_prof_folder,
artifact_name=torch_prof_artifact_name,
num_traces_to_keep=torch_prof_num_traces_to_keep,
overwrite=torch_prof_overwrite,
record_shapes=torch_prof_record_shapes,
profile_memory=torch_prof_profile_memory,
use_gzip=torch_prof_use_gzip,
with_stack=torch_prof_with_stack,
with_flops=torch_prof_with_flops))
# Append the trace handlers at the end, so profilers will log events before the traces are written.
self.state.callbacks.extend(self.state.profiler.trace_handlers)
loggers = list(ensure_tuple(loggers))
if any(isinstance(x, ProgressBarLogger) for x in loggers):
warnings.warn(
DeprecationWarning(
(f"Specifying the {type(ProgressBarLogger).__name__} via `loggers` is deprecated. Instead, "
"please specify `progress_bar`, `log_to_console`, `log_level`, and `stream` arguments when "
"constructing the trainer. If specified, these arguments will be ignored, as the "
f"{type(ProgressBarLogger).__name__} was already created.")))
else:
loggers.append(
ProgressBarLogger(
progress_bar=progress_bar,
log_to_console=log_to_console,
console_log_level=console_log_level,
stream=console_stream,
))
self.logger = Logger(state=self.state, destinations=loggers, run_name=run_name)
self.state.callbacks = list(cast(List[Callback], loggers)) + self.state.callbacks
self._checkpoint_saver = None
if save_folder is not None:
self._checkpoint_saver = CheckpointSaver(
folder=save_folder,
filename=save_filename,
artifact_name=save_artifact_name,
latest_filename=save_latest_filename,
latest_artifact_name=save_latest_artifact_name,
overwrite=save_overwrite,
weights_only=save_weights_only,
save_interval=save_interval,
num_checkpoints_to_keep=save_num_checkpoints_to_keep,
)
self.state.callbacks.append(self._checkpoint_saver)
self.engine = Engine(
state=self.state,
logger=self.logger,
)
self._validate_every_n_batches = validate_every_n_batches
self._validate_every_n_epochs = validate_every_n_epochs
self._grad_clip_norm = grad_clip_norm
if deterministic_mode:
reproducibility.configure_deterministic_mode()
if compute_training_metrics:
warnings.warn(('Computing model evaluation metrics during training doubles the number of forward passes '
'and may lead to a throughput degradation.'))
train_metrics = model.metrics(train=True)
if isinstance(train_metrics, Metric):
# Forcing metrics to be a MetricCollection simplifies logging results
train_metrics = MetricCollection([train_metrics])
self.train_metrics = self._ensure_metrics_device_and_dtype(train_metrics)
else:
self.train_metrics = None
self.engine.run_event(Event.INIT)
assert isinstance(self.state.model, ComposerModel)
self._original_model = self.state.model # TODO(ravi) -- update the state to add an original model helper
# place the state, model in the proper devices, and initialize from a checkpoint if provided
if self.deepspeed_enabled:
try:
import deepspeed
except ImportError as e:
raise MissingConditionalImportError(extra_deps_group="deepspeed",
conda_package="deepspeed>=0.5.5") from e
assert self._deepspeed_config is not None
self._deepspeed_config = _parse_deepspeed_config(self._deepspeed_config,
state=self.state,
grad_clip_norm=self._grad_clip_norm)
optimizer = ensure_tuple(self.state.optimizers)[0]
(self.state.model, self.state.optimizers, _, _) = deepspeed.initialize(
config=self._deepspeed_config,
model=self.state.model,
optimizer=optimizer,
)
# The deepspeed engine is responsible for serializing the model and optimizer state,
# so these attributes should not be serialized with the composer state.
if "model" in self.state.serialized_attributes:
self.state.serialized_attributes.remove("model")
if "optimizers" in self.state.serialized_attributes:
self.state.serialized_attributes.remove("optimizers")
# If using DeepSpeed, the model must be loaded from checkpoint after the engine has been
# initialized, but if using PyTorch DDP, the model must be loaded before it is wrapped with
# DDP.
self._rng_state = None
if load_path is not None:
self._rng_state = load_checkpoint(state=self.state,
path=load_path,
object_store=load_object_store,
load_weights_only=load_weights_only,
strict_model_weights=load_strict,
chunk_size=load_chunk_size,
progress_bar=load_progress_bar)
reproducibility.seed_all(self.state.seed)
if not self.deepspeed_enabled:
host_model_params = self.state.model.parameters()
self.state.model = self._device.module_to_device(self.state.model)
device_model_params = self.state.model.parameters()
# use surgery to update the parameters of the optimizers, now that the model is on the device
# see https://pytorch.org/docs/stable/optim.html#constructing-it
module_surgery.replace_params_in_optimizer(old_params=host_model_params,
new_params=device_model_params,
optimizers=self.state.optimizers)
# Move any remaining optimizer parameters onto the device
self.state.optimizers = map_collection(self.state.optimizers, self._device.optimizer_to_device)
if dist.get_world_size() > 1:
# Only wrap the module if required
self.state.model = _prepare_ddp_module(self.state.model, self._find_unused_parameters)
@property
def deepspeed_enabled(self) -> bool:
"""``True`` if DeepSpeed is being used for training and ``False`` otherwise.
.. seealso:: `DeepSpeed's documentation <https://www.deepspeed.ai/docs/config-json/>`_
"""
return self._deepspeed_config is not None
@property
def saved_checkpoints(self) -> List[Tuple[Timestamp, List[pathlib.Path]]]:
"""The checkpoint timestamps and filepaths.
This list contains tuples of the save timestamp and the checkpoint filepaths.
This list will have at most ``save_num_checkpoints_to_keep`` entries. The latest checkpoint
will be at the end.
.. note::
When using DeepSpeed, the index of a filepath in each list corresponds to the global rank of
the process that wrote that file. Each filepath is valid only on the process's (rank's) node.
Otherwise, when not using DeepSpeed, each sub-list will contain only one filepath since only rank zero
saves checkpoints.
"""
if self._checkpoint_saver is None:
return []
return self._checkpoint_saver.saved_checkpoints
[docs] def fit(self):
"""Train and evaluate the model on the provided data."""
try:
self._train_loop()
finally:
self.engine.close()
def _ensure_metrics_device_and_dtype(self, metrics: MetricCollection):
# Safety check to ensure the metric and data are on the same device. Normally not
# needed because the metric is automatically on the same device as the model.
# See https://torchmetrics.readthedocs.io/en/latest/pages/overview.html for details.
metrics = self._device.module_to_device(metrics)
# HACK: DeepSpeed somehow manages to convert metric internal states to its own dtype. When
# running with FP16, this tends to result in overflows. Let's assume FP32 is good enough.
for _, metric in metrics.items():
metric.set_dtype(torch.float32) # type: ignore
return metrics
def _compute_and_log_metrics(self, dataloader_label: str, log_level: LogLevel, metrics: MetricCollection):
"""Computes metrics, logs the results, and updates the state.
Args:
dataloader_label (str): The dataloader label.
metrics (MetricCollection): The metrics to compute.
log_level (LogLevel): The LogLevel for logging metrics.
"""
computed_metrics = metrics.compute()
self.logger.data(
log_level=log_level,
data={f'metrics/{dataloader_label}/{name}': val for (name, val) in computed_metrics.items()},
)
self.state.current_metrics[dataloader_label] = computed_metrics
def _spin_dataloaders(self):
"""Spin the dataloaders to restore sampler state.
Only one batch must be loaded to seed the sampler's generator. since only the first batch is being loaded, the
dataloader may not be completely iterated through.
"""
# spin the evaluator dataloaders once to initialize its sampler deterministically
# so it does not affect any other RNG reads
for evaluator in self.state.evaluators:
dataloader = evaluator.dataloader.dataloader
# FFCV dataloaders use their own sampling strategy
if isinstance(dataloader, torch.utils.data.DataLoader) and isinstance(dataloader.sampler,
torch.utils.data.DistributedSampler):
dataloader.sampler.set_epoch(0)
for _ in dataloader:
break
# spin the train dataloader's sampler to get to the state of the desired epoch
for epoch in range(int(self.state.timer.epoch)):
# TODO: hasattr check will be removed while fixing https://github.com/mosaicml/composer/issues/424
if hasattr(self.state.train_dataloader, "sampler") and isinstance(self.state.train_dataloader.sampler,
torch.utils.data.DistributedSampler):
self.state.train_dataloader.sampler.set_epoch(epoch)
for _ in self.state.train_dataloader:
break
def _train_loop(self) -> None:
"""Run training for the specified number of epochs and log results."""
# print training start
self.logger.data_fit({"trainer/algorithms": [str(algo) for algo in self.state.algorithms]})
self.engine.run_event(Event.FIT_START)
self.state.scaler = ClosureGradScaler() if self._use_closures() else GradScaler()
use_grad_scaling = self._use_grad_scaling(self.state.precision, self.state.scaler)
self._spin_dataloaders()
if self.state.timer.batch_in_epoch == 0 and self._rng_state is not None:
# only restore the rng state here if the step in the current epoch is zero.
reproducibility.load_rng_state(self._rng_state)
self._rng_state = None
while self.state.timer < self.state.max_duration:
try:
self.state.model.train()
if int(self.state.timer.batch_in_epoch) == 0:
self.engine.run_event(Event.EPOCH_START)
self.logger.data_epoch({"epoch": int(self.state.timer.epoch)})
if self.train_metrics is not None:
# reset the metrics before every epoch
self.train_metrics.reset()
# TODO: hasattr check will be removed while fixing https://github.com/mosaicml/composer/issues/424
if hasattr(self.state.train_dataloader, "sampler") and isinstance(self.state.train_dataloader.sampler,
torch.utils.data.DistributedSampler):
self.state.train_dataloader.sampler.set_epoch(int(self.state.timer.epoch))
for batch_idx, self.state.batch in enumerate(
itertools.islice(self.state.train_dataloader, self.state.steps_per_epoch)):
# if resuming, skip dataloader forward to the minibatch index
if batch_idx < int(self.state.timer.batch_in_epoch):
# Restore the RNG state immediately before the next batch is yielded from the dataloader
if batch_idx + 1 == int(self.state.timer.batch_in_epoch) and self._rng_state is not None:
reproducibility.load_rng_state(self._rng_state)
self._rng_state = None
continue
self.state.batch = self._device.batch_to_device(self.state.batch)
self.state.batch = self._train_data_spec.device_transforms(self.state.batch)
self.state.batch_num_samples = self._train_data_spec.get_num_samples_in_batch(self.state.batch)
self.state.batch_num_tokens = self._train_data_spec.get_num_tokens_in_batch(self.state.batch)
if self.deepspeed_enabled:
self.state.batch = _fix_batch_precision_for_deepspeed(self.state.batch, self.state.precision)
if self.train_metrics is not None:
self.state.model.eval()
with torch.no_grad():
for eval_microbatch in self._train_data_spec.split_batch(
self.state.batch, self.state.grad_accum):
# TODO: Detect if self.run_event(Event.AFTER_DATALOADER) changes the training
# data and if so print a warning that metrics may return unexpected results
self.train_metrics.update(*self._original_model.validate(eval_microbatch))
self.state.model.train()
self.engine.run_event(Event.AFTER_DATALOADER)
num_samples_in_batch = self._device.tensor_to_device(
torch.tensor([self.state.batch_num_samples], dtype=torch.int))
num_tokens_in_batch = self._device.tensor_to_device(
torch.tensor([self.state.batch_num_tokens], dtype=torch.int))
dist.all_reduce(num_samples_in_batch, reduce_operation="SUM")
dist.all_reduce(num_tokens_in_batch, reduce_operation="SUM")
self.engine.run_event(Event.BATCH_START)
self.logger.data_batch({
"trainer/global_step": int(self.state.timer.batch),
"trainer/batch_idx": self.state.timer.batch_in_epoch.value,
})
total_loss = self._train_batch(use_grad_scaling)
if use_grad_scaling:
self.state.scaler.update()
if total_loss is not None:
if not isinstance(total_loss, torch.Tensor):
total_loss = self._device.tensor_to_device(torch.tensor([total_loss]))
# total_loss can be None if gradient scaling failed
dist.all_reduce(total_loss, reduce_operation="SUM")
full_loss = total_loss.cpu().item()
self.logger.data_batch({'loss/train': full_loss / dist.get_world_size()})
self.state.timer.on_batch_complete(
samples=int(num_samples_in_batch.item()),
tokens=int(num_tokens_in_batch.item()),
)
if self._step_schedulers_every_batch:
for scheduler in self.state.schedulers:
scheduler.step()
if self.train_metrics is not None:
self._compute_and_log_metrics(
dataloader_label='train',
log_level=LogLevel.BATCH,
metrics=self.train_metrics,
)
self.engine.run_event(Event.BATCH_END)
if self._validate_every_n_batches > 0 and int(
self.state.timer.batch) % self._validate_every_n_batches == 0:
self.eval(log_level=LogLevel.BATCH)
self.engine.run_event(Event.BATCH_CHECKPOINT)
if self.state.timer >= self.state.max_duration:
# If max_duration is specified in batches, samples, or tokens, and
# and the max_duration is reached mid-epoch, then break out of the dataloader
# to finish the epoch early and finish training.
break
except BreakEpochException:
log.info(f'Skipping the rest of Epoch {int(self.state.timer.epoch)}')
self.state.timer.on_epoch_complete()
if self.train_metrics is not None:
self._compute_and_log_metrics(
dataloader_label='train',
log_level=LogLevel.EPOCH,
metrics=self.train_metrics,
)
if not self._step_schedulers_every_batch:
for scheduler in self.state.schedulers:
scheduler.step()
self.engine.run_event(Event.EPOCH_END)
if self._validate_every_n_epochs > 0 and int(self.state.timer.epoch) % self._validate_every_n_epochs == 0:
self.eval(log_level=LogLevel.EPOCH)
self.engine.run_event(Event.EPOCH_CHECKPOINT)
def _is_cuda_oom(self, e: RuntimeError):
"""Determines if error is CUDA Out of Memory and if adaptive_grad_accum is enabled."""
return self.adaptive_grad_accum and "CUDA out of memory" in str(e)
def _handle_cuda_oom(self):
"""Handles CUDA Out of Memory and rescales if using adaptive grad_accum."""
# Raise runtime error if training 1 sample at a time still resulted in CUDA out of memory
if self.state.grad_accum == self.state.batch_num_samples:
raise RuntimeError(
textwrap.dedent("""CUDA out of memory. Train loop failed with an internal microbatch of size 1.
This means the GPU does not have enough memory to process even 1 sample."""))
else:
self.state.grad_accum = min(2 * self.state.grad_accum, self.state.batch_num_samples)
self.logger.data_batch({'trainer/grad_accum': self.state.grad_accum})
def _train_batch(self, use_grad_scaling: bool):
"""Compute loss by training on a full batch of data. Adaptively change microbatch size if enabled to maximize
GPU usage.
Args:
use_grad_scaling (bool): Enables gradient scaling
"""
# Retry until we successfully complete training and return loss
while True:
total_loss = None
# Note: We use uint8 instead of bool as BOR is not supported on all torch.distributed backends
should_handle_cuda_oom = 0
caught_timeout_error = None
try:
assert self.state.scaler is not None
microbatches = self._train_data_spec.split_batch(self.state.batch, self.state.grad_accum)
if self.deepspeed_enabled:
total_loss = self._train_microbatches(microbatches)
elif self._use_closures():
for optimizer in self.state.optimizers:
if use_grad_scaling:
total_loss = self.state.scaler.step(
optimizer, closure=lambda **kwargs: self._train_microbatches(microbatches, **kwargs))
else:
total_loss = optimizer.step(
closure=lambda **kwargs: self._train_microbatches(microbatches, **kwargs).item())
else:
total_loss = self._train_microbatches(microbatches)
for optimizer in self.state.optimizers:
if use_grad_scaling:
self.state.scaler.step(optimizer)
else:
optimizer.step()
except RuntimeError as e:
if self._is_cuda_oom(e):
log.debug(
textwrap.dedent(f"""Rank {dist.get_global_rank()} OOM'd.
grad_accum will be increased prior to reattempting training on the current batch."""))
should_handle_cuda_oom = 1
elif "Timed out" in str(e):
# Catch timeout errors and only reraise if we did not encounter OOM on other ranks. Error
# is likely transient if one rank OOMed, it likely did not reach a barrier. Note that if we
# catch non-transient timeout errors they will be later reraised if no rank OOMed.
caught_timeout_error = e
else:
raise
# Propagate across all ranks if any rank hit CUDA OOM
should_handle_cuda_oom = self._device.tensor_to_device(
torch.tensor([should_handle_cuda_oom], dtype=torch.uint8))
dist.all_reduce(should_handle_cuda_oom, reduce_operation="MAX")
if int(should_handle_cuda_oom.item()) == 1:
# If any rank hit CUDA OOM, update grad_accum and retry. Ignore any caught_timeout_error since
# it is likely transient, e.g. timeout because certain ranks OOMed and didn't reach barrier.
self._handle_cuda_oom()
elif caught_timeout_error:
# If not CUDA out of memory, raise exception to user. Note that this truncates the call stack
# back only to this newly raised error.
raise caught_timeout_error
else:
# Otherwise, return calculated loss
return total_loss
def _train_microbatches(self, microbatches: Sequence[Batch], ddp_sync: bool = True):
"""Iterate over microbatches and compute the loss that will be used to step the optimizer.
Args:
microbatches (Sequence[Batch]): The microbatches which make up the batch.
ddp_sync (bool): True to sync gradients between devices on every backwards
pass and False to only sync gradients after each device has finished
computing a gradient on it's entire set of microbatches. (default: ``True``)
"""
if ddp_sync or not isinstance(self.state.model, DistributedDataParallel):
context = contextlib.nullcontext
else:
context = cast(Callable[[], ContextManager], self.state.model.no_sync)
with context():
self.engine.run_event(Event.BEFORE_TRAIN_BATCH)
assert self.state.optimizers is not None
assert self.state.scaler is not None
use_grad_scaling = self._use_grad_scaling(self.state.precision, self.state.scaler)
if not self.deepspeed_enabled:
for optimizer in self.state.optimizers:
optimizer.zero_grad()
# tracker for gradient accumulation
total_loss = self._device.tensor_to_device(torch.zeros(size=(1,)))
current_batch_size = sum([self._train_data_spec.get_num_samples_in_batch(batch) for batch in microbatches])
for microbatch_idx, self.state.batch in enumerate(microbatches):
is_final_microbatch = microbatch_idx + 1 == len(microbatches)
self._train_microbatch(use_grad_scaling, current_batch_size, total_loss, is_final_microbatch)
# Unscale gradients before `Event.AFTER_TRAIN_BATCH`
if use_grad_scaling:
for optimizer in ensure_tuple(self.state.optimizers):
self.state.scaler.unscale_(optimizer)
# clip gradients if the magnitude is too large
if not self.deepspeed_enabled and self._grad_clip_norm is not None:
torch.nn.utils.clip_grad_norm_(
parameters=self.state.model.parameters(),
max_norm=self._grad_clip_norm,
)
self.engine.run_event(Event.AFTER_TRAIN_BATCH)
return total_loss
def _train_microbatch(self, use_grad_scaling: bool, current_batch_size: int, total_loss: torch.Tensor,
is_final_microbatch: bool):
"""Train and compute the loss of ``state.batch``, which is assumed to be a single microbatch.
Args:
use_grad_scaling (bool): Whether to use gradient scaling.
minibatch_num_samples (int): Number of samples in the minibatch.
total_loss (torch.Tensor): Total loss aggregated across all microbatches.
is_final_microbatch (bool): If current microbatch is the last one.
"""
assert self.state.scaler is not None
microbatch_num_samples = self._train_data_spec.get_num_samples_in_batch(self.state.batch)
sync_context = contextlib.nullcontext() if self.deepspeed_enabled else _ddp_sync_context(
self.state, is_final_microbatch, self._ddp_sync_strategy)
with sync_context:
# forward pass
self.engine.run_event(Event.BEFORE_FORWARD)
with self.state.precision_context:
self.state.outputs = self.state.model(self.state.batch)
self.engine.run_event(Event.AFTER_FORWARD)
# loss
self.engine.run_event(Event.BEFORE_LOSS)
with self.state.precision_context:
self.state.loss = self._original_model.loss(self.state.outputs, self.state.batch)
# We always want to scale loss by the grad_accum before the backwards pass and
# also for sake of metrics. Complicating matters, the DeepSpeed engine does its
# own scaling when we call `.backward`, but this isn't in place so we still need
# to scale for sake of metrics after the `.backward` call.
# Loss is added to losses with clone to not scale the loss for the step printout
# Likely need to look into the performance impact
if not self.deepspeed_enabled:
for loss in ensure_tuple(self.state.loss):
loss.mul_(microbatch_num_samples / current_batch_size)
total_loss += loss.detach().clone()
assert self.state.loss is not None
self.engine.run_event(Event.AFTER_LOSS)
# backward
self.engine.run_event(Event.BEFORE_BACKWARD)
if use_grad_scaling:
self.state.loss = cast(torch.Tensor, self.state.scaler.scale(self.state.loss))
if self.deepspeed_enabled:
self.state.deepspeed_model.backward(self.state.loss)
# This is the same loss scaling and reporting we skipped earlier.
for loss in ensure_tuple(self.state.loss):
loss.mul_(microbatch_num_samples / current_batch_size)
total_loss += loss.detach().clone()
else:
for loss in ensure_tuple(self.state.loss):
loss.backward(create_graph=self._backwards_create_graph)
self.engine.run_event(Event.AFTER_BACKWARD)
if self.deepspeed_enabled:
self.state.deepspeed_model.step()
[docs] def eval(self, log_level: LogLevel = LogLevel.FIT):
"""Evaluate the model on the provided evaluation data and log appropriate metrics.
Args:
log_level (LogLevel, optional): The log level to use for metric logging during evaluation.
Defaults to :attr:`~.LogLevel.FIT`.
"""
restore_model_train = self.state.model.training
self.state.model.eval()
with torch.no_grad():
self.engine.run_event(Event.EVAL_START)
for evaluator in self.state.evaluators:
dataloader = evaluator.dataloader.dataloader
metrics = self._ensure_metrics_device_and_dtype(evaluator.metrics)
metrics.reset()
# TODO: hasattr check will be removed while fixing https://github.com/mosaicml/composer/issues/424
if hasattr(dataloader, "sampler") and isinstance(dataloader.sampler,
torch.utils.data.DistributedSampler):
# The distributed sampler uses `set_epoch` to set the random seed
# Because evaluation can run on each batch, we use the batch to seed the sampler
# so each evaluation will get a proper shuffle.
# The epoch provided to `set_epoch` need not be sequential, so this is fine.
dataloader.sampler.set_epoch(int(self.state.timer.batch))
for self.state.batch in itertools.islice(dataloader, self._eval_subset_num_batches):
self.state.batch = self._device.batch_to_device(self.state.batch)
if evaluator.dataloader.device_transforms:
self.state.batch = evaluator.dataloader.device_transforms(self.state.batch)
self.state.batch_num_samples = evaluator.dataloader.get_num_samples_in_batch(self.state.batch)
self.state.batch_num_tokens = evaluator.dataloader.get_num_tokens_in_batch(self.state.batch)
if self.deepspeed_enabled:
self.state.batch = _fix_batch_precision_for_deepspeed(self.state.batch, self.state.precision)
self.engine.run_event(Event.EVAL_BATCH_START)
self.engine.run_event(Event.EVAL_BEFORE_FORWARD)
self.state.outputs, targets = self._original_model.validate(self.state.batch)
self.engine.run_event(Event.EVAL_AFTER_FORWARD)
metrics.update(self.state.outputs, targets)
self._compute_and_log_metrics(dataloader_label=evaluator.label,
metrics=metrics,
log_level=log_level)
self.engine.run_event(Event.EVAL_BATCH_END)
self.logger.data_epoch({"epoch": self.state.timer.epoch.value})
self.logger.data_batch({"trainer/global_step": self.state.timer.batch.value})
self._compute_and_log_metrics(dataloader_label=evaluator.label, metrics=metrics, log_level=log_level)
self.engine.run_event(Event.EVAL_END)
if restore_model_train:
self.state.model.train()
def _use_grad_scaling(self, precision: Union[str, Precision], scaler: Optional[GradScaler]) -> bool:
"""Determines based on precision when to use grad scaling.
By default, the pytorch GradScaler is a no-op if running on
unsupported hardware. Here we raise a RuntimeError instead.
Args:
precision (Precision): Numerical precision, based on the Precision Enum.
scaler (GradScaler): Used to make sure that the scaler is enabled when
using grad scaling.
Raises:
RuntimeError:
Occurs when attempting to use grad scaling without the scaler
enabled. Likely due to hardware not supporting the provided precision.
"""
if self.deepspeed_enabled:
return False
precision = Precision(precision)
use_grad_scaling = precision == Precision.AMP
if use_grad_scaling and (scaler is None or not scaler.is_enabled()):
raise RuntimeError(f'Attempting to use grad scaling with {precision}, but scaler is not enabled.'
f'Potentially your hardware does not support Precision {precision}.')
return use_grad_scaling
def _use_closures(self) -> bool:
"""Determines based on precision and optimizers whether to use closures.
We default to using closures unless AMP is enabled, in which case we only allow closures when using optimizers
with the _step_supports_amp_closure flag.
"""
if self.deepspeed_enabled:
return False
if self.state.precision != Precision.AMP:
return True
if self.state.optimizers is None:
raise RuntimeError("state.optimizers must be set before `_use_closures` can be determined")
return all(
getattr(optimizer, "_step_supports_amp_closure", False)
for optimizer in ensure_tuple(self.state.optimizers))
[docs] def save_checkpoint(self, name: str = "ep{epoch}-ba{batch}-rank{rank}", *, weights_only: bool = False):
"""Checkpoint the training :class:`~.State`.
Args:
name (str, optional): See :func:`.save_checkpoint`.
weights_only (bool, optional): See :func:`.save_checkpoint`.
Returns:
List[pathlib.Path]: See :func:`.save_checkpoint`.
"""
return save_checkpoint(state=self.state, logger=self.logger, filename=name, weights_only=weights_only)