# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0
"""Train models."""
from __future__ import annotations
import collections.abc
import contextlib
import datetime
import itertools
import logging
import os
import random
import re
import tempfile
import textwrap
import time
import warnings
from collections import defaultdict
from copy import deepcopy
from pathlib import Path
from typing import (
Any,
Callable,
ContextManager,
Dict,
Iterable,
List,
Mapping,
Optional,
Sequence,
TextIO,
Tuple,
Union,
cast,
)
import coolname
import torch
import torch.distributed
import torch.nn as nn
import torch.utils.data
from packaging import version
from torch._dynamo import OptimizedModule
from torch.cuda.amp.grad_scaler import GradScaler
from torch.distributed.fsdp import FullyShardedDataParallel
from torch.distributed.fsdp._runtime_utils import _post_backward_final_callback
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
from torch.nn.parallel import DistributedDataParallel
from torch.optim.lr_scheduler import LRScheduler
from torch.utils.data import DataLoader, DistributedSampler
from torchmetrics import Metric
if version.parse(torch.__version__) >= version.parse('2.3.0'):
from torch.amp.grad_scaler import GradScaler, _refresh_per_optimizer_state # type: ignore
else:
from torch.cuda.amp.grad_scaler import GradScaler, _refresh_per_optimizer_state # type: ignore
from composer.callbacks import CheckpointSaver, MemorySnapshot, OOMObserver, OptimizerMonitor
from composer.core import (
Algorithm,
AlgorithmPass,
Batch,
Callback,
DataSpec,
Engine,
Evaluator,
Event,
Precision,
State,
Time,
Timestamp,
TimeUnit,
TrainerMode,
ensure_data_spec,
ensure_evaluator,
ensure_time,
get_precision_context,
)
from composer.devices import Device, DeviceCPU, DeviceGPU, DeviceMPS, DeviceTPU
from composer.loggers import (
ConsoleLogger,
Logger,
LoggerDestination,
MLFlowLogger,
MosaicMLLogger,
ProgressBarLogger,
RemoteUploaderDownloader,
WandBLogger,
)
from composer.loggers.mosaicml_logger import MOSAICML_ACCESS_TOKEN_ENV_VAR, MOSAICML_PLATFORM_ENV_VAR
from composer.models import ComposerModel
from composer.optim import ComposerScheduler, DecoupledSGDW, compile_composer_scheduler
from composer.profiler import Profiler
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.dist_strategy import (
DDPSyncStrategy,
ddp_sync_context,
prepare_ddp_module,
prepare_fsdp_module,
set_fsdp_default,
)
from composer.utils import (
ExportFormat,
MissingConditionalImportError,
ObjectStore,
Transform,
checkpoint,
dist,
ensure_tuple,
export_with_logger,
extract_hparams,
format_name_with_dist,
get_composer_env_dict,
get_device,
get_file,
is_xla_installed,
map_collection,
maybe_create_object_store_from_uri,
maybe_create_remote_uploader_downloader_from_uri,
model_eval_mode,
parse_uri,
partial_format,
reproducibility,
)
from composer.utils.misc import is_model_deepspeed
from composer.utils.object_store.mlflow_object_store import MLFLOW_EXPERIMENT_ID_FORMAT_KEY, MLFLOW_RUN_ID_FORMAT_KEY
if is_xla_installed():
import torch_xla.core.xla_model as xm
import torch_xla.distributed.parallel_loader as pl
log = logging.getLogger(__name__)
__all__ = ['Trainer']
# syntax to shorten the Scheduler type annotations
Scheduler = Union[ComposerScheduler, LRScheduler]
def _raise_missing_argument_exception(arg_name: str):
raise ValueError((
f'{arg_name} is a required argument and must be specified when constructing the '
f'{Trainer.__name__} or when calling {Trainer.__name__}.{Trainer.fit.__name__}(). '
f'To fix, please specify `{arg_name}` via {Trainer.__name__}({arg_name}=...) or '
f'{Trainer.__name__}.{Trainer.fit.__name__}({arg_name}=...).'
))
def _scale_max_duration_by_ssr(
scale_schedule_ratio: float,
orig_max_duration: Optional[Time[int]],
) -> Optional[Time[int]]:
if orig_max_duration is None:
return None
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.')
return max_duration
def _get_default_scheduler_frequency(schedulers: Optional[Union[Scheduler, Sequence[Scheduler]]]):
has_pytorch_scheduler = any(isinstance(scheduler, LRScheduler) for scheduler in ensure_tuple(schedulers))
if has_pytorch_scheduler:
log.info((
'Stepping schedulers every epoch, as a PyTorch scheduler was provided. '
'The trainer cannot automatically convert the parameters (e.g. step_size, T_max) of the '
'PyTorch scheduler to be in terms of batches. If the PyTorch scheduler should be stepped '
'every batch, set `step_schedulers_every_batch=True`.'
))
return TimeUnit.EPOCH
else:
log.info((
'Stepping schedulers every batch. '
'To step schedulers every epoch, set `step_schedulers_every_batch=False`.'
))
return TimeUnit.BATCH
def _filter_metrics(metrics: Dict[str, Metric], metric_names: Optional[List[str]]) -> Dict[str, Metric]:
"""Filter the metrics based on the given metric_names as regex strings (e.g. 'Accuracy', 'f1' for 'BinaryF1Score', 'Top-.' for 'Top-1 Accuracy' and 'Top-2 Accuracy', etc). If no metric_names are provided, all metrics will be returned."""
metrics = deepcopy(metrics)
if metric_names is None:
return metrics
filtered_metrics = {}
for name, metric in metrics.items():
if any(re.match(f'.*{metric_name}.*', name, re.IGNORECASE) for metric_name in metric_names):
filtered_metrics[name] = metric
return filtered_metrics
def _validate_precision(precision: Precision, device: Device):
if isinstance(device, DeviceCPU) and precision != Precision.FP32:
raise ValueError(f'{precision} is not supported for CPU training.')
def _compile_schedulers(
schedulers: Optional[Union[Scheduler, Sequence[Scheduler]]],
state: State,
scale_schedule_ratio: float,
) -> List[LRScheduler]:
compiled_schedulers = []
for scheduler in ensure_tuple(schedulers):
if isinstance(scheduler, LRScheduler):
scale_pytorch_scheduler(scheduler, scale_schedule_ratio)
compiled_schedulers.append(scheduler)
# It's a composer scheduler
else:
compiled_schedulers.append(
compile_composer_scheduler(
scheduler,
state,
scale_schedule_ratio,
),
)
return compiled_schedulers
def _set_evaluator_interval_and_subset_num_batches(
evaluators: Sequence[Evaluator],
eval_interval: Union[int, str, Time, Callable[[State, Event], bool]],
subset_num_batches: int,
):
# Convert eval_dataloader to `List[Evaluator]`
for evaluator in evaluators:
if evaluator.subset_num_batches is None:
evaluator.subset_num_batches = subset_num_batches
if evaluator.eval_interval is None:
evaluator.eval_interval = eval_interval
eval_dataloader = evaluator.dataloader.dataloader
if isinstance(eval_dataloader, collections.abc.Sized) and evaluator.subset_num_batches == -1:
try:
dataloader_len = len(eval_dataloader)
except TypeError:
dataloader_len = None
if dataloader_len == None:
raise ValueError(
'eval_subset_num_batches must be set when using an infinite sized '
'eval_dataloader where length is `None`. Otherwise, evaluation will '
'run forever and never terminate.',
)
def _is_auto_microbatching(device_train_microbatch_size: Optional[Union[int, float, str]], device: Device):
if device_train_microbatch_size == 'auto':
warnings.warn((
"`device_train_microbatch_size='auto'` may potentially fail with unexpected "
'CUDA errors. Auto microbatching attempts to catch CUDA Out of Memory errors '
'and adjust the batch size, but it is possible CUDA will be put into an '
'irrecoverable state due to PyTorch bugs, e.g. integer overflow. In this case, '
'please manually set device_train_microbatch_size explicitly to an integer '
'instead.'
))
if not isinstance(device, DeviceGPU):
raise ValueError(
'Can only use adaptive device_train_microbatch_size on GPU. Please set device_train_microbatch_size >= 1.',
)
return True
else:
return False
def _get_initial_device_train_microbatch_size(
device_train_microbatch_size: Optional[Union[int, float, str]],
auto_microbatching: bool,
train_dataloader: Optional[Iterable],
) -> Optional[Union[int, float]]:
"""Sets initial value of device_train_microbatch_size.
If auto_microbatching, sets initial `device_train_microbatch_size` to per rank batch size. If
`train_dataloader` is not set yet, returns None and this function will be called again when
`train_dataloader` is set, such as when `fit()` is called.
"""
if device_train_microbatch_size is None or auto_microbatching:
# Return None, this function will be called again when `train_dataloader` is set
if train_dataloader is None:
return None
try:
batch_size = getattr(train_dataloader, 'batch_size')
except AttributeError as e:
# Error message when `device_train_microbatch_size` is None
# Note: This code path will be removed after `auto` is made default
if device_train_microbatch_size is None:
raise ValueError(
'`device_train_microbatch_size` must be set when `state.train_dataloader` does not have a `batch_size` attribute.',
) from e
# Error message when `device_train_microbatch_size` is 'auto'
raise AttributeError(
"`device_train_microbatch_size='auto'` requires the `state.train_dataloader` to have a `batch_size` attribute.",
) from e
return batch_size
elif isinstance(device_train_microbatch_size, Union[int, float]):
return device_train_microbatch_size
else:
raise ValueError("device_train_microbatch_size must be an int or ``'auto'``")
def _is_cuda_oom(e: RuntimeError):
"""Determines if error is CUDA Out of Memory and if auto_microbatching is enabled."""
if 'CUDA out of memory' in str(e):
return True
# With batch_norm, large batch sizes sometimes result in cuDNN instead of Cuda OOMs.
if 'cuDNN error: CUDNN_STATUS_NOT_SUPPORTED. This error may appear if you passed in a non-contiguous input.' in str(
e,
):
warnings.warn(
'Encountered "cuDNN error: CUDNN_STATUS_NOT_SUPPORTED. This error may appear if you passed in '
'a non-contiguous input." This can happen when the batch_size is too large for the GPU so auto '
'auto_microbatching will rerun with a smaller microbatch size value, but there may be a user '
'error with non-contiguous inputs.',
)
return True
return False
def _fsdp_reshard_and_cleanup(model: torch.nn.Module):
"""Manually reshard and clean up FSDP model.
When an exception like OOM happens, _post_backward_final_callback, which
is registered as a backward callback, will not run. We manually call it to cleanup
loose memory.
"""
for __, module in model.named_modules():
if isinstance(module, FullyShardedDataParallel):
if module.check_is_root():
# Only call _post_backward_final_callback on root module. It will
# traverse and reshard all FSDP sub-modules
_post_backward_final_callback(module, module)
def _adjust_device_train_microbatch_size(state: State):
"""Adjust device_train_microbatch_size if we encounter OOM.
Args:
state (State): State of trainer.
"""
# If any rank hit CUDA OOM, update device_train_microbatch_size and retry. Raise runtime error
# if training 1 sample at a time still resulted in CUDA out of memory.
assert state.device_train_microbatch_size is not None
if state.device_train_microbatch_size == 1:
raise RuntimeError((
'CUDA out of memory. The train loop failed with an internal microbatch of size 1.'
'The GPU does not have enough memory to process even 1 sample during train.'
))
else:
original_microbatch_size = state.device_train_microbatch_size
state.device_train_microbatch_size = max(int(original_microbatch_size / 2), 1)
warnings.warn(
RuntimeWarning(
'CUDA out of memory detected. Train microbatch size will be decreased from '
f'{original_microbatch_size} -> {state.device_train_microbatch_size}.',
),
)
# Clear gradients in case failure happened during backwards pass
if hasattr(state, 'outputs'):
del state.outputs
if hasattr(state, 'loss'):
del state.loss
for optimizer in state.optimizers:
optimizer.zero_grad(set_to_none=True)
if state.scaler is not None:
state.scaler._per_optimizer_states = defaultdict(_refresh_per_optimizer_state)
_fsdp_reshard_and_cleanup(state.model)
torch.cuda.empty_cache()
def _adjust_device_eval_microbatch_size(evaluator: Evaluator):
"""Adjust device_eval_microbatch_size if we encounter OOM.
Args:
evaluator (State): Current evaluator
"""
# If any rank hit CUDA OOM, update device_eval_microbatch_size and retry. Raise runtime error
# if evaluating 1 sample at a time still resulted in CUDA out of memory.
assert evaluator.device_eval_microbatch_size is not None
if evaluator.device_eval_microbatch_size == 1:
raise RuntimeError((
'CUDA out of memory. The eval loop failed with an internal microbatch of size 1.'
'The GPU does not have enough memory to process even 1 sample during eval.'
))
else:
original_microbatch_size = evaluator.device_eval_microbatch_size
evaluator.device_eval_microbatch_size = max(int(original_microbatch_size / 2), 1)
warnings.warn(
RuntimeWarning(
'CUDA out of memory detected. Train microbatch size will be decreased from '
f'{original_microbatch_size} -> {evaluator.device_eval_microbatch_size}.',
),
)
torch.cuda.empty_cache()
def _validate_evaluator(evaluator: Evaluator, device: Device):
"""Ensure automicrobatching is only on GPU.
Unlike `device_train_microbatch_size`, this validation must be done separately from the
`_is_auto_microbatching` check because `device` is not available during `Evaluator`
initialization.
"""
auto_microbatching = evaluator.auto_microbatching
if auto_microbatching and not isinstance(device, DeviceGPU):
raise ValueError(
'Can only use adaptive device_eval_microbatch_size on GPU. Please set device_eval_microbatch_size >= 1.',
)
if evaluator.auto_microbatching and hasattr(evaluator.dataloader, 'seq_parallel_world_size'):
raise ValueError(
'Auto microbatching on evaluators is not compatible with sequence parallelism. '
'Please manually set device_eval_microbatch_size or disable sequence parallelism .',
)
if hasattr(
evaluator.dataloader,
'seq_parallel_world_size',
) and evaluator.dataloader.seq_parallel_world_size > 1 and abs( # type: ignore
evaluator.device_eval_microbatch_size * evaluator.dataloader.seq_parallel_world_size - 1, # type: ignore
) > 1e-4:
raise ValueError(
'Sequence parallelism requires a microbatch size of 1 distributed over the sequence parallel group.',
)
def _distribute_and_get_random_seed(seed: Optional[int], device: Device):
if seed is None:
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)')
# using int64 to prevent overflow
rank_zero_seed = device.tensor_to_device(torch.tensor([seed], dtype=torch.int64))
if dist.get_world_size() > 1:
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()
return rank_zero_seed, seed
def _get_ddp_sync_strategy(ddp_sync_strategy: Optional[Union[str, DDPSyncStrategy]], find_unused_parameters: bool):
if ddp_sync_strategy is None:
if find_unused_parameters:
ddp_sync_strategy = DDPSyncStrategy.MULTI_AUTO_SYNC
else:
ddp_sync_strategy = DDPSyncStrategy.SINGLE_AUTO_SYNC
else:
ddp_sync_strategy = DDPSyncStrategy(ddp_sync_strategy)
return ddp_sync_strategy
def _get_precision_context(precision: Precision, precision_config: Optional[Dict[str, Any]], deepspeed_enabled: bool):
if deepspeed_enabled:
return contextlib.nullcontext()
return get_precision_context(precision, precision_config)
def _generate_run_name() -> str:
# change coolname randomness for different names with same seed
coolname.replace_random(random.Random(os.urandom(128)))
# prefixing with the time so experiments sorted alphabetically will have the latest experiment last
generated_run_name = str(int(time.time())) + '-' + coolname.generate_slug(2)
run_name_list = [generated_run_name]
# ensure all ranks have the same experiment name
dist.broadcast_object_list(run_name_list)
generated_run_name = run_name_list[0]
return generated_run_name
def _get_distributed_sampler(dataloader: DataLoader) -> Optional[DistributedSampler]:
"""Fetch a distributed sampler from a `dataloader` if it exists."""
if isinstance(dataloader.batch_sampler, DistributedSampler):
return dataloader.batch_sampler
if isinstance(dataloader.sampler, DistributedSampler):
return dataloader.sampler
return None
[docs]class Trainer:
"""Train models with Composer algorithms.
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.
See the :doc:`training guide</trainer/using_the_trainer>` for more information.
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",
eval_interval="1ep",
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()
# 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",
eval_interval="1ep",
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()
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 (Iterable | DataSpec | dict, optional): The 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 dataloader.
It is recommended that the dataloader, whether specified directly or as part of a :class:`.DataSpec`,
should be a :class:`torch.utils.data.DataLoader`.
.. note:: The ``train_dataloader`` should yield per-rank batches. Each per-rank batch
will then be further divided based on the ``device_train_microbatch_size`` 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 ``device_train_microbatch_size = 128``,
then the per-rank batch will be divided into ``256 / 128 = 2`` microbatches of size ``128``.
If ``train_dataloader`` is not specified when constructing the trainer, it must be specified when invoking
:meth:`.Trainer.fit`.
train_dataloader_label (str, optional): The label for the train dataloader. (default: ``'train'``)
This label is used to index the training metrics in
:attr:`.State.train_metrics`.
This parameter has no effect if ``train_dataloader`` is not specified.
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 ``-1``, then the entire dataloader will be iterated over. (default: ``-1``)
When using the profiler, it can be helpful to set this parameter to the length of the profile schedule.
This setting will end each epoch early to avoid additional training that will not be profiled.
This parameter is ignored if ``train_dataloader`` is not specified.
spin_dataloaders (bool, optional): If ``True``, dataloaders will be spun up to the current timestamp
by skipping samples which have already been trained on. If a dataloader has a way to resume from
the current batch without spinning, this will be a no-op. This ensures dataloaders continue from
the same batch when resuming training. (default: ``True``)
.. note:: Spinning dataloaders can be potentially very slow but is required to skip samples which
have already been trained on. If it is acceptable to repeat samples when resuming training,
it is possible to resume faster by setting ``spin_dataloaders=False``. This may have severe
performance implications and is generally not recommended unless you confidently understand the
implications.
max_duration (Time | str | int, optional): 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.
If ``max_duration`` is not specified when constructing the trainer, ``duration`` must be specified when invoking
:meth:`.Trainer.fit`.
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.
algorithm_passes ([AlgorithmPass | Tuple[AlgorithmPass, int] | Sequence[AlgorithmPass | Tuple[AlgorithmPass, int]], optional):
Optional list of passes to change order in which algorithms are applied. These passes are merged with the
default passes specified in :class:`.Engine`. If ``None``, then no additional passes will be used.
(default: ``None``)
.. seealso:: :class:`composer.core.Engine` for more information.
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 (LRScheduler | ComposerScheduler | Sequence[LRScheduler | 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.
scale_schedule_ratio (float, optional): Ratio by which to scale the training duration and learning rate
schedules. (default: ``1.0``)
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.
This parameter has no effect if ``schedulers`` is not specified.
.. 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``)
eval_dataloader (Iterable | DataLoader | DataSpec | Evaluator | Sequence[Evaluator], optional): The :class:`.Iterable`,
: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:`.DataLoader`, :class:`.DataSpec`, or :class:`.Iterable` is passed in, then all
metrics returned by ``model.get_metrics()`` will be used during evaluation. If a :class:`.Evaluator`
is specified in a list, all eval dataloaders must be :class:`.Evaluator` instances.
``None`` results in no evaluation. (default: ``None``)
eval_interval (int | str | Time | (State, Event) -> bool, optional): Specifies how frequently to run evaluation.
An integer, which will be interpreted to be epochs, a str (e.g. ``1ep``, or ``10ba``), a :class:`.Time`
object, or a callable.
Defaults to ``1`` (evaluate every epoch).
If an integer (in epochs), :class:`.Time` string, or :class:`.Time` instance, the evaluator will be run
with this frequency. :class:`.Time` strings or :class:`.Time` instances must have units of
:attr:`.TimeUnit.BATCH` or :attr:`.TimeUnit.EPOCH`.
Set to ``0`` to disable evaluation.
If a callable, it should take two arguments (:class:`.State`, :class:`.Event`) and return a bool
representing whether the evaluator should be invoked. The event will be either :attr:`.Event.BATCH_END`
or :attr:`.Event.EPOCH_END`.
This ``eval_interval`` will apply to any :class:`.Evaluator` in ``eval_dataloader`` that does not specify
an ``eval_interval`` or if a dataloader is passed in directly. This parameter has no effect if
``eval_dataloader`` is not specified.
When specifying time string or integer for the ``eval_interval``, the evaluator(s) are also run at the ``Event.FIT_END`` if it doesn't
evenly divide the training duration.
eval_subset_num_batches (int, optional): If specified, evaluate on this many batches. Defaults to ``-1``,
which means to iterate over the entire dataloader.
This parameter has no effect if ``eval_dataloader`` is not specified, it is greater than
``len(eval_dataloader)``, or ``eval_dataloader`` is an :class:`.Evaluator` and ``subset_num_batches``
was specified as part of the :class:`.Evaluator`.
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.
loggers (LoggerDestination | Sequence[LoggerDestination], optional): The destinations to log training information to.
.. seealso:: :mod:`composer.loggers` for the different loggers built into Composer.
run_name (str, optional): A name for this training run. If not specified, the env var
`COMPOSER_RUN_NAME` or `RUN_NAME` will be used if set. Otherwise, the timestamp will be
combined with a :doc:`coolname <coolname:index>`, e.g. ``1654298855-electric-zebra``.
progress_bar (bool): Whether to show a progress bar. (default: ``True``)
log_to_console (bool): Whether to print logging statements to the console. (default: ``False``)
console_stream (TextIO | str, optional): The stream to write to. If a string, it can either be
``'stdout'`` or ``'stderr'``. (default: :attr:`sys.stderr`)
console_log_interval (int | str | Time, optional): Specifies how frequently to log metrics to console.
An integer, which will be interpreted to be epochs, a str (e.g. ``1ep``, or ``10ba``), a :class:`.Time`
object, or a callable. (default: ``1ba``)
Defaults to ``1ba`` (log metrics every batch).
If an integer (in epochs), :class:`.Time` string, or :class:`.Time` instance, the metrics will be logged
with this frequency. :class:`.Time` strings or :class:`.Time` instances must have units of
:attr:`.TimeUnit.BATCH` or :attr:`.TimeUnit.EPOCH`.
Set to ``0`` to disable metrics logging to console.
log_traces (bool): Whether to log traces or not. (default: ``False``)
auto_log_hparams (bool): Whether to automatically extract hyperparameters. (default: ``False``)
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. If a URI is specified, ``load_object_store`` does not need to be set.
When using `Deepspeed ZeRO <https://www.deepspeed.ai/tutorials/zero/>`_, checkpoints are sharded 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 (Union[ObjectStore, LoggerDestination], optional): If the ``load_path`` is in an
object store (i.e. AWS S3 or Google Cloud Storage), an instance of :class:`.ObjectStore` or
:class:`.LoggerDestination` which will be used to retrieve the checkpoint. Otherwise, if the
checkpoint is a local filepath, set to ``None``. Also, it can be ``None`` if the ``load_path`` is
an S3 URI because the appropriate object store will be automatically constructed in that case.
Ignored if ``load_path`` is ``None``.
(default: ``None``)
Example:
.. testsetup::
import composer.trainer
composer.trainer.trainer.checkpoint.load_checkpoint = lambda *args, **kwargs: None
.. testcode::
from composer import Trainer
from composer.utils import LibcloudObjectStore
# Create the object store provider with the specified credentials
creds = {"key": "object_store_key",
"secret": "object_store_secret"}
store = LibcloudObjectStore(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",
eval_interval="1ep",
load_path=checkpoint_path,
load_object_store=store,
)
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_model_weights (bool, optional): Ensure that the set of weights in the checkpoint and model must exactly match.
Ignored if ``load_path`` is ``None``. (default: ``True``)
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``)
load_ignore_keys (List[str] | (Dict) -> None, optional): A list of paths for the ``state_dict`` of the checkpoint,
which, when provided, will be ignored from the state_dict before a checkpoint is loaded. Each path is a list
of strings specifying the keys to index into ``state_dict`` joined together with `/` as a separator (as PyTorch
uses `.` in parameter names). If a prefix is provided, all children are also ignored (see Example 2).
See :mod:`composer.core.state` for the structure of state_dict.
Example 1: ``load_ignore_keys = ["state/model/layer1.weights", "state/model/layer1.bias"]`` would ignore
layer 1 weights and bias.
Example 2: ``load_ignore_keys = ["state/model/*"]`` would ignore the entire model, which would have the same
effect as the previous example if there was only 1 layer.
Example 3: ``load_ignore_keys = ["state/model/layer*.weights"]`` would ignore all weights in the model.
Example 4: ``load_ignore_keys = ["state/rank_zero_seed", "rng"]`` would reset all randomness when
loading the checkpoint.
If a callable, it should take one argument which is the state_dict. The callable is free to arbitrarily modify
the state_dict before it is loaded.
(default: ``None``)
load_exclude_algorithms (List[str], optional): A list of algorithm names to exclude from loading.
By default, algorithms with `required_on_load=True` which were enabled when training the loaded
checkpoint are automatically applied unless they conflict with a user specified algorithm. These
algorithms often change the model, and not applying them could result in certain layers not having
weights loaded.
Example 1: ``load_exclude_algorithms = ["BlurPool"]`` would exclude BlurPool from loading.
Example 2: ``load_exclude_algorithms = ["FusedLayerNorm", "Alibi"]`` would exclude FusedLayerNorm and Alibi from loading.
(default: ``None``)
save_folder (str, optional): Format string for the folder where checkpoints are saved.
If ``None``, checkpoints will not be saved. Can also be a URI for S3 paths only.
In the case of an S3 URI, the appropriate `~.RemoteUploader` object will be created
automatically. (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}.pt"``)
.. 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}.pt"``)
.. 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_ignore_keys (List[str] | (Dict) -> None, optional): A list of paths for the ``state_dict`` of the checkpoint,
which, when provided, will be ignored from the state_dict before a checkpoint is saved. Each path is a list
of strings specifying the keys to index into ``state_dict`` joined together with `/` as a separator (as PyTorch
uses `.` in parameter names). If a prefix is provided, all children are also ignored (see Example 2).
See :mod:`composer.core.state` for the structure of state_dict.
Example 1: ``save_ignore_keys = ["state/model/layer1.weights", "state/model/layer1.bias"]`` would ignore
layer 1 weights and bias.
Example 2: ``save_ignore_keys = ["state/model/*"]`` would ignore the entire model, which would have the same
effect as the previous example if there was only 1 layer.
Example 3: ``save_ignore_keys = ["state/model/layer*.weights"]`` would ignore all weights in the model.
Example 4: ``save_ignore_keys = ["state/rank_zero_seed", "rng"]`` would reset all randomness when
saving the checkpoint.
If a callable, it should take one argument which is the state_dict. The callable is free to arbitrarily modify
the state_dict before it is loaded.
(default: ``None``)
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 uploaded. For example, when this callback
is used in conjunction with the :class:`.RemoteUploaderDownloader`, 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
remote file systems.
save_metrics (bool, optional): Whether to save the metrics. By default, metrics are not saved to checkpoint
as state usually does not need to be preserved and inconsistent state can cause issues when loading.
(default: ``False``)
autoresume (bool, optional): Whether or not to enable autoresume, which allows for stopping and resuming
training. This allows use of spot instances, as the training run is now fault tolerant. This parameter
requires ``save_folder`` and ``run_name`` to be specified and ``save_overwrite`` to be ``False``.
(default: ``False``)
When enabled, the save_folder is checked for checkpoints of the format ``"{save_folder}/{save_latest_filename}"``,
which are loaded to continue training. If no local checkpoints are found, each logger is checked for potential
remote checkpoints named ``"{save_folder}/{save_latest_filename}"``. Finally, if no logged checkpoints are found, ``load_path`` is
used to load a checkpoint if specified. This should only occur at the start of a run using autoresume.
For example, to run a fine-tuning run on a spot instance, ``load_path`` would be set to the original
weights and an object store logger would be added. In the original run, ``load_path`` would be used
to get the starting checkpoint. For any future restarts, such as due to the spot instance being killed,
the loggers would be queried for the latest checkpoint the object store logger would be downloaded and
used to resume training.
deepspeed_config (Dict[str, Any], optional): Configuration for DeepSpeed, formatted as a JSON
according to `DeepSpeed's documentation <https://www.deepspeed.ai/docs/config-json/>`_. (default: ``None``)
To use DeepSpeed with default values, set to the empty dictionary ``{}``.
To disable DeepSpeed (the default), set to ``None``.
fsdp_config (Dict[str, Any], optional): Configuration for FSDP.
See :doc:`FSDP Documentation </notes/distributed_training>` for more details.
To use FSDP with default values, set to the empty dictionary ``{}``. To
disable FSDP, set to ``None``. (default: ``None``)
fsdp_auto_wrap (bool, optional): option to let trainer wrap the module, or if
the module is already wrapped outside, allow the user to disable auto-wrapping.
device (Device | str, optional): The device to use for training, which can be ``'cpu'``, ``'gpu'``,
``'tpu'``, or ``'mps'``. (default: ``None``)
The default behavior sets the device to ``'gpu'`` if CUDA is available, and otherwise ``'cpu'``.
precision (Precision | str, optional): Numerical precision to use for training. One of ``fp32``, ``amp_bf16``
or ``amp_fp16`` (recommended). (default: ``Precision.FP32`` if training on CPU; ``Precision.AMP_FP16`` if
training on GPU)
precision_config (Optional[Dict[str, Any]]): The config for FP8 scaling strategy. See parameters for
`DelayedScaling <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/api/common.html?highlight=delayedscaling#transformer_engine.common.recipe.DelayedScaling>`_.
device_train_microbatch_size (Union[int, float, str), optional): The number of samples to process on each device per
microbatch during training. Gradients are summed over the microbatches per device. If set to ``auto``,
dynamically decreases device_train_microbatch_size if microbatch is too large for GPU. (default: ``None``)
.. note:: This is implemented by taking the batch yielded by the ``train_dataloader`` and splitting
it into sections of size ``device_train_microbatch_size``. If the batch size of the dataloader
is not divisible by ``device_train_microbatch_size``, the last section will be potentially smaller.
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.
dist_timeout (float, optional): Timeout, in seconds, for initializing the distributed process group.
(default: ``300.0``)
ddp_sync_strategy (str | DDPSyncStrategy, optional): The strategy to use for synchronizing gradients.
Leave unset to let the trainer auto-configure this. See :class:`.DDPSyncStrategy`
for more details.
profiler (Profiler, optional): The profiler, if profiling should be enabled. (default: ``None``)
.. seealso::
See the :doc:`Profiling Guide </trainer/performance_tutorials/profiling>` for
additional information.
python_log_level (str, optional): The Python log level to use for log statements in the :mod:`composer`
module. (default: ``None``). If it is ``None``, python logging will not be configured (i.e.
``logging.basicConfig`` won't be called).
.. seealso:: The :mod:`logging` module in Python.
compile_config (Dict[str, Any], optional): Configuration for torch compile. Only supported with PyTorch 2.0 or higher.
Checkout [`torch.compile`](https://pytorch.org/get-started/pytorch-2.0/) for more details.
To use torch compile with default values, set it to empty dictionary ``{}``.
To use torch compile with custom config, set to a dictionary such as ``{'mode': 'max-autotune'}``.
To disable torch compile, set to ``None``. (default: ``None``)
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,
*,
# The Model
model: ComposerModel,
# Train Dataloader
train_dataloader: Optional[Union[Iterable, DataSpec, Dict[str, Any]]] = None,
train_dataloader_label: str = 'train',
train_subset_num_batches: int = -1,
spin_dataloaders: bool = True,
# Stopping Condition
max_duration: Optional[Union[int, str, Time]] = None,
# Algorithms
algorithms: Optional[Union[Algorithm, Sequence[Algorithm]]] = None,
# Engine Pass Registration
algorithm_passes: Optional[Union[AlgorithmPass,
Tuple[AlgorithmPass, int],
Sequence[Union[AlgorithmPass, Tuple[AlgorithmPass, int]]],
]] = None,
# Optimizers and Scheduling
optimizers: Optional[torch.optim.Optimizer] = None,
schedulers: Optional[Union[ComposerScheduler,
LRScheduler,
Sequence[Union[ComposerScheduler,
LRScheduler,
]],
]] = None,
scale_schedule_ratio: float = 1.0,
step_schedulers_every_batch: Optional[bool] = None,
# Evaluators
eval_dataloader: Optional[Union[Iterable, DataSpec, Evaluator, Sequence[Evaluator]]] = None,
eval_interval: Union[int, str, Time, Callable[[State, Event], bool]] = 1,
eval_subset_num_batches: int = -1,
# Callbacks and Logging
callbacks: Optional[Union[Callback, Sequence[Callback]]] = None,
loggers: Optional[Union[LoggerDestination, Sequence[LoggerDestination]]] = None,
run_name: Optional[str] = None,
progress_bar: bool = True,
log_to_console: bool = False,
console_stream: Union[str, TextIO] = 'stderr',
console_log_interval: Union[int, str, Time] = '1ba',
log_traces: bool = False,
auto_log_hparams: bool = False,
# Load Checkpoint
load_path: Optional[str] = None,
load_object_store: Optional[Union[ObjectStore, LoggerDestination]] = None,
load_weights_only: bool = False,
load_strict_model_weights: bool = True,
load_progress_bar: bool = True,
load_ignore_keys: Optional[Union[List[str], Callable[[Dict], None]]] = None,
load_exclude_algorithms: Optional[List[str]] = None,
# Save Checkpoint
save_folder: Optional[str] = None,
save_filename: str = 'ep{epoch}-ba{batch}-rank{rank}.pt',
save_latest_filename: Optional[str] = 'latest-rank{rank}.pt',
save_overwrite: bool = False,
save_interval: Union[str, int, Time, Callable[[State, Event], bool]] = '1ep',
save_weights_only: bool = False,
save_ignore_keys: Optional[Union[List[str], Callable[[Dict], None]]] = None,
save_num_checkpoints_to_keep: int = -1,
save_metrics: bool = False,
# Graceful Resumption
autoresume: bool = False,
# DeepSpeed
deepspeed_config: Optional[Dict[str, Any]] = None,
fsdp_config: Optional[Dict[str, Any]] = None,
fsdp_auto_wrap: bool = True,
# System/Numerics
device: Optional[Union[str, Device]] = None,
precision: Optional[Union[str, Precision]] = None,
precision_config: Optional[Dict[str, Any]] = None,
device_train_microbatch_size: Optional[Union[int, float, str]] = None,
# Reproducibility
seed: Optional[int] = None,
deterministic_mode: bool = False,
# Distributed Training
dist_timeout: float = 300.0,
ddp_sync_strategy: Optional[Union[str, DDPSyncStrategy]] = None,
# Profiling
profiler: Optional[Profiler] = None,
# Python logging
python_log_level: Optional[str] = None,
# compile config for PyTorch 2.0 or higher
compile_config: Optional[Dict[str, Any]] = None,
):
self.auto_log_hparams = auto_log_hparams
self.python_log_level = python_log_level
if self.python_log_level is not None:
logging.basicConfig(
# Example of format string
# 2022-06-29 11:22:26,152: rank0[822018][MainThread]: INFO: composer.trainer.trainer: Using precision Precision.FP32
# Including the PID and thread name to help with debugging dataloader workers and callbacks that spawn background
# threads / processes
format=
f'%(asctime)s: rank{dist.get_global_rank()}[%(process)d][%(threadName)s]: %(levelname)s: %(name)s: %(message)s',
)
logging.getLogger('composer').setLevel(self.python_log_level.upper())
algorithms = list(ensure_tuple(algorithms))
# Device
device = get_device(device)
# Precision
if precision is None:
precision = Precision.AMP_FP16 if isinstance(device, DeviceGPU) else Precision.FP32
elif isinstance(precision, str):
precision = Precision(precision)
_validate_precision(precision, device)
# check if provided model is compiled or not
is_model_compiled = False
if isinstance(model, OptimizedModule):
log.warning(
f'Provided `model` is already compiled with `torch.compile`. Ignoring ' +
f'parameter `compile_config` if provided. If you would like `Trainer` ' +
f'to takes care of model compilation, provide a not-compiled model and ' +
f'`compile_config` parameter.',
)
# The `torch.compile` function returns an object of type `torch._dynamo.OptimizedModule`
# which wraps the original `nn.Module` object and later patches its forward method to
# optimized `self.forward` method.
is_model_compiled = True
compiled_model = model._orig_mod
if not isinstance(compiled_model, ComposerModel):
raise ValueError(
f'Provided `model` must be a subclass of ComposerModel. ' +
f'Instead found as type `{type(compiled_model)}`',
)
compiled_model.forward = model.dynamo_ctx(compiled_model.forward)
model = compiled_model
# Microbatching
auto_microbatching = _is_auto_microbatching(device_train_microbatch_size, device=device)
if auto_microbatching and train_dataloader is not None and hasattr(train_dataloader, 'seq_parallel_world_size'):
raise ValueError('`device_train_microbatch_size="auto"` is not compatible with sequence parallelism.')
if train_dataloader is not None and hasattr(
train_dataloader,
'seq_parallel_world_size',
) and train_dataloader.seq_parallel_world_size > 1 and abs( # type: ignore
device_train_microbatch_size * train_dataloader.seq_parallel_world_size - 1, # type: ignore
) > 1e-4:
raise ValueError(
'`Sequence parallelism requires a microbatch size of 1 distributed over the sequence parallel group.',
)
if auto_microbatching and profiler:
raise ValueError(
"`device_train_microbatch_size='auto'` is not compatible with the profiler. It is "
"recommended to run a mini-run with `device_train_microbatch_size='auto'` to identify "
'the optimal device_train_microbatch_size value and then manually specify that in a '
'second run with profiler.',
)
self.first_batch_complete = False
# If auto_microbatching is True or `device_train_microbatch_size` is not specified, the microbatch size
# will be determined when dataloader is specified. train_dataloader is parsed after `Event.INIT` or in
# fit()
device_train_microbatch_size = _get_initial_device_train_microbatch_size(
device_train_microbatch_size,
auto_microbatching,
None,
)
assert not isinstance(device_train_microbatch_size, str)
# Distributed
if deepspeed_config is not None or fsdp_config is not None or dist.get_world_size() > 1:
# Deepspeed and FSDP both require torch.distributed to be initialized, even if the world size is 1
# And torch.distributed is always required for multi-rank training
dist.initialize_dist(device, dist_timeout)
# Reproducibility
rank_zero_seed, seed = _distribute_and_get_random_seed(seed, device)
# 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)
if deterministic_mode:
reproducibility.configure_deterministic_mode()
# Optimizers and Schedulers
if optimizers is None:
try:
optimizers = DecoupledSGDW(model.parameters(), lr=0.1)
# hard-coding the optimizer in the warning, as repr(optimizers) would print an annoying, multi-line warning
warnings.warn((
'No optimizer was specified. Defaulting to '
f"{type(optimizers).__name__}(lr={optimizers.defaults['lr']})"
))
except ValueError as e:
if 'optimizer got an empty parameter list' in str(e):
warnings.warn(
'No optimizer was specified, and the model does not have parameters. Skipping auto-creating optimizer.',
)
else:
raise
if optimizers is not None:
num_optimizers = len(ensure_tuple(optimizers))
if num_optimizers != 1:
raise NotImplementedError(f'Only one optimizer is supported; found {num_optimizers} optimizers')
# Move the model and optimizers to the device
if deepspeed_config is None and fsdp_config is None:
# check if model is already on tpu
if isinstance(device, DeviceTPU) and 'xla' not in str(next(model.parameters()).device):
raise ValueError(
'Use model.to(xm.xla_device()) to set the model to the TPU before providing to the trainer.',
)
else:
model = device.module_to_device(model)
# Move any remaining optimizer parameters onto the device
# It is possible that optimizer initialize created some internal tensors on CPU
# that need to be moved onto GPU.
optimizers = map_collection(optimizers, device.optimizer_to_device)
# Run Name
run_name = os.getenv('COMPOSER_RUN_NAME', None) if run_name is None else run_name
run_name = os.getenv('RUN_NAME', None) if run_name is None else run_name
if run_name is None:
if autoresume:
raise ValueError('When autoresume=True, the `run_name` must be specified.')
run_name = _generate_run_name()
log.info('Run name: %s', run_name)
# Create the State
self.state = State(
rank_zero_seed=rank_zero_seed,
algorithms=algorithms,
model=model,
device=device,
callbacks=callbacks,
device_train_microbatch_size=device_train_microbatch_size,
auto_microbatching=auto_microbatching,
precision=precision,
precision_config=precision_config,
optimizers=optimizers,
run_name=run_name,
save_metrics=save_metrics,
deepspeed_config=deepspeed_config,
fsdp_config=set_fsdp_default(fsdp_config) if fsdp_config is not None else None,
fsdp_auto_wrap=fsdp_auto_wrap,
)
# Console Logging
loggers = list(ensure_tuple(loggers))
# Profiler
if profiler is not None:
warnings.warn('The profiler is enabled. Using the profiler adds additional overhead when training.')
self.state.profiler = profiler
for remote_uri in profiler.remote_filenames:
remote_ud = maybe_create_remote_uploader_downloader_from_uri(uri=remote_uri, loggers=loggers)
if remote_ud is not None:
loggers.append(remote_ud)
self.state.profiler.bind_to_state(self.state)
# MemorySnapshot, OOMObserver
for cb in self.state.callbacks:
if isinstance(cb, MemorySnapshot) or isinstance(cb, OOMObserver):
if cb.remote_file_name:
remote_ud = maybe_create_remote_uploader_downloader_from_uri(
uri=cb.remote_file_name,
loggers=loggers,
)
if remote_ud is not None:
loggers.append(remote_ud)
if progress_bar and log_to_console:
warnings.warn(
'Setting both `progress_bar` and `log_to_console` both to True is not recommended and will'
'lead to duplicate logs and weird formatting issues. Please set one of them to False for a better logging experience.',
)
if any(isinstance(x, ProgressBarLogger) for x in loggers):
warnings.warn(
Warning((
f'Specifying the {ProgressBarLogger.__name__} via `loggers` is not recommended as '
'any values set for the following Trainer arguments will be ignored: `progress_bar`, `console_stream`, or `log_traces`. '
'The recommended way of enabling a progress bar is to set `progress_bar` to True instead of '
f'constructing a {ProgressBarLogger.__name__} instance.'
)),
)
else:
if progress_bar:
loggers.append(ProgressBarLogger(stream=console_stream, log_traces=log_traces))
# Console Logging
if any(isinstance(x, ConsoleLogger) for x in loggers):
warnings.warn(
Warning((
f'Specifying the {ConsoleLogger.__name__} via `loggers` is not recommended as '
'any values set for the following Trainer arguments will be ignored: `log_to_console`, `console_stream`, `log_traces`, and `console_log_interval`. '
'The recommended way of enabling a console logging is to set `log_to_console` to True instead of '
f'constructing a {ConsoleLogger.__name__} instance.'
)),
)
else:
if log_to_console:
loggers.append(
ConsoleLogger(stream=console_stream, log_interval=console_log_interval, log_traces=log_traces),
)
# MosaicML Logger
# Keep MosaicML logger above the RemoteUploaderDownloader so that fit end is reported before the final checkpoint begins uploading
if os.environ.get(MOSAICML_PLATFORM_ENV_VAR, 'false').lower() == 'true' and os.environ.get(
MOSAICML_ACCESS_TOKEN_ENV_VAR,
) is not None and not any(isinstance(x, MosaicMLLogger) for x in loggers):
log.info('Detected run on MosaicML platform. Adding MosaicMLLogger to loggers.')
mosaicml_logger = MosaicMLLogger()
loggers.append(mosaicml_logger)
# Remote Uploader Downloader
# Keep the ``RemoteUploaderDownloader`` below client-provided loggers so the loggers init callbacks run before
# the ``RemoteUploaderDownloader`` init. This is necessary to use an ``MLFlowObjectStore`` to log objects to a
# run managed by an ``MLFlowLogger``, as the ``MLFlowObjectStore`` relies on the ``MLFlowLogger`` to initialize
# the active MLFlow run.
if save_folder is not None:
remote_ud = maybe_create_remote_uploader_downloader_from_uri(save_folder, loggers)
if remote_ud is not None:
loggers.append(remote_ud)
# Logger
self.logger = Logger(state=self.state, destinations=loggers)
if save_latest_filename is not None:
remote_ud_has_format_string = [
isinstance(logger_destination, RemoteUploaderDownloader) and
logger_destination.file_path_format_string != '{remote_file_name}'
for logger_destination in self.logger.destinations
]
if any(remote_ud_has_format_string):
raise ValueError(
'Specifying a `file_path_format_string` to a `RemoteUploaderDownloader` is not currently supported while using `save_latest_filename`. '
'Please specify the path formatting via `save_folder`, `save_filename`, and `save_latest_filename`',
)
# Callbacks
self.state.callbacks[:] = list(cast(List[Callback], loggers)) + self.state.callbacks
# Checkpoint Saving
self._checkpoint_saver = None
latest_remote_file_name = None
if save_folder is not None:
if save_weights_only:
log.info(
'save_weights_only=True now also saves metadata and integrations! Please adjust your workflow accordingly.',
)
_, _, parsed_save_folder = parse_uri(save_folder)
# If user passes a URI with s3:// and a bucket_name, but no other
# path then we assume they just want their checkpoints saved directly in their
# bucket.
if parsed_save_folder == '':
folder = '.'
remote_file_name = save_filename
latest_remote_file_name = save_latest_filename
# If they actually specify a path, then we use that for their local save path
# and we prefix save_filename with that path for remote_file_name.
else:
folder = parsed_save_folder
remote_file_name = str(Path(parsed_save_folder) / Path(save_filename))
if save_latest_filename is not None:
latest_remote_file_name = str(Path(parsed_save_folder) / Path(save_latest_filename))
else:
latest_remote_file_name = None
self._checkpoint_saver = CheckpointSaver(
folder=folder,
filename=save_filename,
remote_file_name=remote_file_name,
latest_filename=save_latest_filename,
latest_remote_file_name=latest_remote_file_name,
overwrite=save_overwrite,
weights_only=save_weights_only,
ignore_keys=save_ignore_keys,
save_interval=save_interval,
num_checkpoints_to_keep=save_num_checkpoints_to_keep,
)
self.state.callbacks.append(self._checkpoint_saver)
# The Engine
self.engine = Engine(state=self.state, logger=self.logger, algorithm_passes=algorithm_passes)
# Set the logger
self.state.model.logger = self.logger # pyright: ignore[reportGeneralTypeIssues]
# Run Event.INIT
self.engine.run_event(Event.INIT)
# If the experiment is being tracked with an `MLFlowLogger`, then MLFlow experiment and run are available
# after Event.INIT.
if save_folder is not None:
mlflow_logger = None
for destination in self.logger.destinations:
if isinstance(destination, MLFlowLogger):
mlflow_logger = destination
break
if mlflow_logger is not None:
mlflow_experiment_id = mlflow_logger._experiment_id
mlflow_run_id = mlflow_logger._run_id
# The save folder and related paths/filenames may contain format placeholders for the MLFlow IDs, so
# populate them now.
mlflow_format_kwargs = {
MLFLOW_EXPERIMENT_ID_FORMAT_KEY: mlflow_experiment_id,
MLFLOW_RUN_ID_FORMAT_KEY: mlflow_run_id,
}
save_folder = partial_format(save_folder, **mlflow_format_kwargs)
if latest_remote_file_name is not None:
latest_remote_file_name = partial_format(latest_remote_file_name, **mlflow_format_kwargs)
# Log hparams.
if self.auto_log_hparams:
locs = locals()
if 'cb' in locs:
del locs['cb']
self.local_hparams = extract_hparams(locs)
self.logger.log_hyperparameters(self.local_hparams)
# Log composer version
composer_env_dict = get_composer_env_dict()
self.logger.log_hyperparameters({'composer_version': composer_env_dict['composer_version']})
self.logger.log_hyperparameters({'composer_commit_hash': str(composer_env_dict['composer_commit_hash'])})
# Log gpus and nodes.
device_name = self.state.device.__class__.__name__.lstrip('Device').lower()
self.logger.log_hyperparameters({
'num_nodes': int(dist.get_world_size() / dist.get_local_world_size()),
f'num_{device_name}s_per_node': dist.get_local_world_size(),
'node_name': os.environ.get('NODENAME', 'unknown because NODENAME environment variable not set'),
})
if not isinstance(self.state.model, ComposerModel):
raise ValueError('Provided model must be a subclass of ComposerModel.')
# After running Event.INIT, then set the "optional" elements of state that could be passed in on FIT instead of INIT
# Setting these attributes here ensures that algorithms do not depend on unavailable attributes during Event.INIT
# Metrics and Evaluators
# Set state.train_metrics and state.eval_metrics here to allow callbacks / algs to potentially
# change the model, which could change what metrics are computed
self.state.train_metrics = deepcopy(self.state.model.get_metrics(is_train=True))
self.state.eval_metrics = {}
if eval_dataloader is None:
evaluators: List[Evaluator] = []
else:
eval_metrics = deepcopy(self.state.model.get_metrics(is_train=False))
model_metric_names = [str(k) for k in eval_metrics.keys()]
eval_dataloader = ensure_tuple(eval_dataloader)
evaluator_types = [isinstance(evaluator, Evaluator) for evaluator in eval_dataloader]
if any(evaluator_types) and not all(evaluator_types):
raise ValueError(
'Mixing Evaluator with other classes is not allowed, please wrap'
'all other classes with the Evaluator class. These are the classes'
'that were detected:' + str([type(evaluator) for evaluator in eval_dataloader]),
)
evaluators = [
ensure_evaluator(evaluator, default_metric_names=model_metric_names) for evaluator in eval_dataloader
]
# match metric names to model metrics
self.state.eval_metrics = {
evaluator.label: _filter_metrics(eval_metrics, evaluator.metric_names) for evaluator in evaluators
}
_set_evaluator_interval_and_subset_num_batches(
evaluators=evaluators,
eval_interval=eval_interval,
subset_num_batches=eval_subset_num_batches,
)
for evaluator in evaluators:
_validate_evaluator(evaluator, self.state.device)
if len(evaluators) == 0:
if eval_subset_num_batches != -1:
warnings.warn(
f'Specifying `eval_subset_num_batches={eval_subset_num_batches}` without an `eval_dataloader` '
'has no effect. If trying to run an evaluator, make sure `eval_dataloader` is specified. '
'Otherwise, set `eval_subset_num_batches` to default value -1.',
)
if eval_interval != 0 and eval_interval != 1:
warnings.warn(
f'Specifying `eval_interval={eval_interval}` without an `eval_dataloader` has no effect. '
'If trying to run an evaluator, make sure `eval_dataloader` is specified. Otherwise, '
'set `eval_interval` to 0 or default value 1.',
)
self.state.evaluators = evaluators
# Train Dataloader
self._train_data_spec = None if train_dataloader is None else ensure_data_spec(train_dataloader)
if self._train_data_spec is not None:
self.state.set_dataloader(
self._train_data_spec.dataloader,
train_dataloader_label,
train_subset_num_batches,
)
if self.state.device.dist_backend == 'xla':
self.state.train_dataloader = pl.MpDeviceLoader(self.state.dataloader, xm.xla_device())
else:
self.state.train_dataloader = self.state.dataloader
self.state.device_train_microbatch_size = _get_initial_device_train_microbatch_size(
self.state.device_train_microbatch_size,
self.state.auto_microbatching,
self.state.train_dataloader,
)
self.spin_dataloaders = spin_dataloaders
# Max Duration
if max_duration is not None:
self.state.max_duration = ensure_time(max_duration, TimeUnit.EPOCH)
self.logger.log_hyperparameters({'rank_zero_seed': rank_zero_seed})
# Schedulers
self.state.schedulers = _compile_schedulers(schedulers, self.state, scale_schedule_ratio)
if scale_schedule_ratio != 1.0:
if len(self.state.schedulers) == 0:
raise ValueError('Specifying `scale_schedule_ratio` without `schedulers` has no effect.')
self.state.max_duration = _scale_max_duration_by_ssr(scale_schedule_ratio, self.state.max_duration)
if step_schedulers_every_batch is None:
self._scheduler_step_frequency = _get_default_scheduler_frequency(schedulers)
else:
self._scheduler_step_frequency = TimeUnit.BATCH if step_schedulers_every_batch else TimeUnit.EPOCH
# Some algorithms require specific settings
self._backwards_create_graph = any((x.backwards_create_graph for x in self.state.algorithms))
self._find_unused_parameters = any((x.find_unused_parameters for x in self.state.algorithms))
self._ddp_sync_strategy = _get_ddp_sync_strategy(ddp_sync_strategy, self._find_unused_parameters)
# Suppressing 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')
self.state.scaler = ClosureGradScaler() if self._use_closures() else GradScaler()
if self.state.fsdp_config is not None:
# This state should never be reached, but we raise a ValueError just in case
if self._use_closures() and self.state.precision == Precision.AMP_FP16:
raise ValueError(
f'Using closures and precision {self.state.precision} is not supported'
f' with FSDP. Please use another optimizer or precision type.',
)
self.state.scaler = ShardedGradScaler()
# suppressing FSDP warning when auto grad accum exits the forward pass before completing
warnings.filterwarnings(action='ignore', message='Forward order differs from that of the first iteration')
# If using DDP or DeepSpeed, we need to wrap the ComposerModel but store a reference to the
# original model for functions like `eval_forward`, `get_metrics`, etc.
self._original_model = self.state.model
# If using PyTorch DDP, the model must be loaded before it is wrapped with DDP.
# If using DeepSpeed, the engine must be initialized before the model is loaded.
# If using FSDP, the model must be wrapped and then loaded unless loading a monolith
# checkpoint on rank 0 only, in which case the model be loaded before it is wrapped.
# FSDP wrap if not using monolith checkpoint on rank 0 only
if self.state.fsdp_config is not None and fsdp_auto_wrap and not self.state.load_fsdp_monolith_rank0_only:
with reproducibility.seed_context(self.state.rank_zero_seed):
prepare_fsdp_module(
model,
optimizers,
self.state.fsdp_config,
precision,
device,
auto_microbatching,
self.state.seed,
)
# Configure Deepspeed
if self.state.deepspeed_config is not None:
for callback in self.state.callbacks:
if isinstance(callback, OptimizerMonitor):
raise ValueError(
'OptimizerMonitor is not supported with DeepSpeed because DeepSpeed clears '
'the gradients before in the last call to .backward see: '
'https://github.com/microsoft/DeepSpeed/issues/2329 for more details.',
)
try:
import deepspeed
except ImportError as e:
raise MissingConditionalImportError(
extra_deps_group='deepspeed',
conda_package='deepspeed>=0.5.5',
conda_channel=None,
) from e
self.state.deepspeed_config = _parse_deepspeed_config(self.state.deepspeed_config, state=self.state)
optimizer = ensure_tuple(self.state.optimizers)[0]
log.debug('Initializing deepspeed')
(self.state.model, self.state.optimizers, _, _) = deepspeed.initialize(
config=self.state.deepspeed_config,
model=self.state.model,
optimizer=optimizer,
)
# Since the DeepSpeed ZeRO optimizer does not inherit torch.optim.Optimizer, the schedulers must be
# compiled and bound BEFORE DeepSpeed initialization. However, this is OK, as the the DeepSpeed Zero
# optimizer uses the same underlying parameter groups as the original optimizer. See
# * https://github.com/microsoft/DeepSpeed/blob/fee73135980e78f8be7e1a3ff556751623ef6aaa/deepspeed/runtime/zero/stage_1_and_2.py#L1911-L1917
# * https://github.com/microsoft/DeepSpeed/blob/ef17c89570ceae5b26a5f886e9d8cd0941afc0ac/deepspeed/runtime/zero/stage3.py#L2532-L2538
# In addition, 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')
self.engine.run_event(Event.BEFORE_LOAD)
# Load Checkpoint
self._rng_state = None
# If autoresume is enabled, first check for existing checkpoints to load
if autoresume:
log.info('Searching for a previous checkpoint to autoresume')
error_message = ''
if save_folder is None:
error_message += 'The `save_folder` must be specified when autoresume is enabled. '
if save_overwrite:
error_message += textwrap.dedent(
'The flag `save_overwrite` must be False when autoresume is enabled as autoresume always loads the '
'latest existing checkpoint in `save_folder`. ',
)
if save_latest_filename is None:
error_message += 'The `save_latest_filename` must be specified so autoresume knows where to load checkpoints from. '
if error_message != '':
raise ValueError(error_message)
assert save_folder is not None
assert save_latest_filename is not None
remote_ud_has_multiple_concurrent_uploads = [
isinstance(logger_destination, RemoteUploaderDownloader) and
logger_destination._num_concurrent_uploads != 1 for logger_destination in self.logger.destinations
]
if any(remote_ud_has_multiple_concurrent_uploads):
raise ValueError(
'Multiple concurrent uploads is not currently supported when using autoresume. Please set `num_concurrent_uploads` to 1 '
'for all `RemoteUploaderDownloader` instances.',
)
assert latest_remote_file_name is not None
if self.state.fsdp_sharded_state_dict_enabled:
ar_object_store = maybe_create_object_store_from_uri(save_folder)
# Symlink is on object store
if ar_object_store is not None:
autoresume_checkpoint_path = None
if dist.get_global_rank() == 0:
with tempfile.TemporaryDirectory() as temp_dir:
local_symlink_file = str(Path(temp_dir) / Path('autoresume.symlink'))
symlink_file_name = format_name_with_dist(
latest_remote_file_name,
self.state.run_name,
) + '.symlink'
try:
ar_object_store.download_object(symlink_file_name, local_symlink_file)
with open(local_symlink_file, 'r') as f:
real_path = f.read()
log.debug(f'Read path {real_path} from symlink file')
autoresume_checkpoint_path = ar_object_store.get_uri(real_path)
except FileNotFoundError:
pass
autoresume_path_list = [autoresume_checkpoint_path]
dist.broadcast_object_list(autoresume_path_list)
autoresume_checkpoint_path = autoresume_path_list[0]
# Symlink is local
else:
save_latest_filename = format_name_with_dist(save_latest_filename, self.state.run_name)
rank0_save_latest_filename = dist.all_gather_object(save_latest_filename)[0]
save_folder = format_name_with_dist(save_folder, self.state.run_name)
latest_checkpoint_path = os.path.join(save_folder, rank0_save_latest_filename)
if os.path.exists(latest_checkpoint_path):
latest_checkpoint_path = os.path.join(
os.path.dirname(latest_checkpoint_path),
os.readlink(latest_checkpoint_path),
)
autoresume_checkpoint_path = latest_checkpoint_path
else:
autoresume_checkpoint_path = None
# Standard non-elastic codepath for autoresume
else:
autoresume_checkpoint_path = self._get_autoresume_checkpoint(
save_folder=save_folder,
save_latest_filename=save_latest_filename,
save_latest_remote_file_name=latest_remote_file_name,
loggers=loggers,
load_progress_bar=load_progress_bar,
)
# Found latest checkpoint path, load that instead
if autoresume_checkpoint_path:
load_path = autoresume_checkpoint_path
# Disable object_store since _get_autoresume_checkpoint will download the checkpoint
# To the save folder, if needed.
load_object_store = None
# Set load arguments to defaults as this applies only to the initial non-autoresume
# load. We do not reset `load_strict_model_weights` for models with frozen layers.
load_weights_only = False
load_ignore_keys = None
load_exclude_algorithms = None
log.info('Autoresuming training from checkpoint')
else:
log.info('No previous autoresume checkpoint found')
# Actually load the checkpoint from potentially updated arguments
if load_path is not None:
if load_object_store is None:
load_object_store = maybe_create_object_store_from_uri(load_path)
if isinstance(load_object_store, WandBLogger):
import wandb
if wandb.run is None:
load_object_store.init(self.state, self.logger)
_, _, parsed_load_path = parse_uri(load_path)
self._rng_state = checkpoint.load_checkpoint(
state=self.state,
logger=self.logger,
path=parsed_load_path,
object_store=load_object_store,
load_weights_only=load_weights_only,
strict_model_weights=load_strict_model_weights,
progress_bar=load_progress_bar,
ignore_keys=load_ignore_keys,
exclude_algorithms=load_exclude_algorithms,
algorithm_passes=self.engine.algorithm_passes,
)
self.state.run_name = run_name
# FSDP wrap if model is not yet wrapped and FSDP is enabled. This can happen if
# load_fsdp_monolith_rank0_only=True but no checkpoint was loaded.
if not self.state.fsdp_enabled and self.state.fsdp_config is not None and self.state.fsdp_auto_wrap and self.state.load_fsdp_monolith_rank0_only:
with reproducibility.seed_context(self.state.rank_zero_seed):
prepare_fsdp_module(model, optimizers, self.state.fsdp_config, precision, device, auto_microbatching)
self.engine.run_event(Event.AFTER_LOAD)
# reseed here. This helps with a couple of issues:
# 1. rng state may change at Event.INIT/Event.BEFORE_LOAD/Event.AFTER_LOAD. For example,
# if an algorithm creates a new module and module parameters are initialized randomly, rng
# state will change. This reseeding nullifies such effects.
# 2. While resuming from a checkpoint, we want to spin dataloader and bring it back to the
# same state as at the time of the checkpoint. Therefore, spinning needs to start from the
# same rng state as in the original run.
log.info(f'Setting seed to {self.state.seed}')
reproducibility.seed_all(self.state.seed)
# DDP wrap if required
if not self.state.deepspeed_enabled and not self.state.fsdp_enabled and dist.get_world_size() > 1:
self.state.model = prepare_ddp_module(self.state.model, self._find_unused_parameters)
# The model would need to be torch.compile()'d after being wrapped in a distributed strategy
# to take advantage of any graph breaks.
if not is_model_compiled and compile_config is not None:
compiled_model = torch.compile(self.state.model, **compile_config)
self.state.model = compiled_model._orig_mod
self.state.model.forward = compiled_model.dynamo_ctx(self.state.model.forward)
is_model_compiled = True
# update local_hparams to ensure the `is_model_compiled` is set correctly for
# debugging purpose and for unit test.
if self.auto_log_hparams:
self.local_hparams['is_model_compiled'] = is_model_compiled
@property
def saved_checkpoints(self) -> List[str]:
"""Returns list of saved checkpoints.
.. note::
For DeepSpeed, which saves file on every rank, only the files corresponding to the process's rank
will be shown.
"""
if self._checkpoint_saver is None:
return []
return self._checkpoint_saver.saved_checkpoints
def _try_checkpoint_download(
self,
latest_checkpoint_path: str,
save_latest_remote_file_name: str,
loggers: Sequence[LoggerDestination],
load_progress_bar: bool,
) -> None:
"""Attempts to download the checkpoint from the logger destinations."""
log.debug(
f'Trying to download {save_latest_remote_file_name} to {latest_checkpoint_path} on rank {dist.get_global_rank()}',
)
for logger in loggers:
try:
# Fetch from logger. If it succeeds, stop trying the rest of the loggers
get_file(
path=save_latest_remote_file_name,
destination=latest_checkpoint_path,
object_store=logger,
overwrite=True,
progress_bar=load_progress_bar,
)
break
except (NotImplementedError, FileNotFoundError):
log.info(f'Checkpoint not found in: {logger}')
# Ignore errors caused by no checkpoint saved with logger
pass
def _get_autoresume_checkpoint(
self,
save_folder: str,
save_latest_filename: str,
save_latest_remote_file_name: str,
loggers: Sequence[LoggerDestination],
load_progress_bar: bool,
) -> Optional[str]:
"""Determines the load path when using autoresume.
First, check the ``save_folder`` for the latest checkpoint.
If no latest checkpoint is found locally, then check each logger for the latest checkpoint, and download
it to the ``save_folder``.
Returns:
Optional[str]: The path to the latest checkpoint, if found, otherwise None.
"""
save_latest_filename = format_name_with_dist(save_latest_filename, self.state.run_name)
save_folder = format_name_with_dist(save_folder, self.state.run_name)
save_latest_remote_file_name = format_name_with_dist(save_latest_remote_file_name, self.state.run_name)
latest_checkpoint_path = os.path.join(save_folder, save_latest_filename)
log.info(
f'Looking for autoresume checkpoint: {save_latest_remote_file_name} (remote), {latest_checkpoint_path} (local)',
)
if self.state.deepspeed_enabled or self.state.fsdp_sharded_state_dict_enabled:
# If latest checkpoint is not saved locally, try to fetch from loggers
if not os.path.exists(latest_checkpoint_path):
log.debug(f'Attempting to download the checkpoint on to rank {dist.get_global_rank()}')
os.makedirs(save_folder, exist_ok=True)
self._try_checkpoint_download(
latest_checkpoint_path,
save_latest_remote_file_name,
loggers,
load_progress_bar,
)
# List of whether the checkpoint exists on each rank
latest_checkpoint_exists = dist.all_gather_object(os.path.exists(latest_checkpoint_path))
if all(latest_checkpoint_exists): # All paths exist, so return the path.
return latest_checkpoint_path
# Require all ranks to have their own local checkpoint if we wish to restore from it for
# deepspeed or fsdp + sharding
elif any(latest_checkpoint_exists): # Some but not all exist, which is very bad.
missing_ranks = [n for (n, exist) in enumerate(latest_checkpoint_exists) if not exist]
mode = 'Deepspeed' if self.state.deepspeed_enabled else 'FSDP sharding'
raise RuntimeError(f'{mode} was enabled, but checkpoints missing on ranks: {missing_ranks}')
else: # None of the paths exists, so no autoresume necessary.
return None
else:
# broadcast the local checkpoint path to all ranks
latest_checkpoint_path_list = [os.path.abspath(latest_checkpoint_path)]
dist.broadcast_object_list(latest_checkpoint_path_list, src=0)
latest_checkpoint_path = latest_checkpoint_path_list[0]
# broadcast the remote checkpoint path to all ranks
save_latest_remote_file_name_list = [save_latest_remote_file_name]
dist.broadcast_object_list(save_latest_remote_file_name_list, src=0)
save_latest_remote_file_name = save_latest_remote_file_name_list[0]
# try to download the checkpoint on local rank 0 of all nodes
if dist.get_local_rank() == 0 and not os.path.exists(latest_checkpoint_path):
log.debug(f'Attempting to download the checkpoint {save_latest_remote_file_name} on to all nodes')
os.makedirs(save_folder, exist_ok=True)
self._try_checkpoint_download(
latest_checkpoint_path,
save_latest_remote_file_name,
loggers,
load_progress_bar,
)
signal_file_path = os.path.join(
os.path.dirname(latest_checkpoint_path),
f'.node_{dist.get_node_rank()}_local_rank0_completed_autoresume',
)
if dist.get_local_rank() == 0:
os.makedirs(os.path.dirname(signal_file_path), exist_ok=True)
with open(signal_file_path, 'wb') as f:
f.write(b'local_rank0_completed_autoresume')
# Avoid the collective call until the local rank zero has finished trying to download the checkpoint
# so that we don't timeout for large downloads. This syncs all processes on the node
with dist.local_rank_zero_download_and_wait(signal_file_path):
# Then, wait to ensure every node has finished downloading the checkpoint
dist.barrier()
if dist.get_local_rank() == 0:
os.remove(signal_file_path)
dist.barrier()
# At this point the rank 0 filepath should exist on all ranks if the download succeeded
# list of whether the checkpoint exists on each rank
latest_checkpoint_exists = dist.all_gather_object(os.path.exists(latest_checkpoint_path))
log.debug(
f'Checkpoint {latest_checkpoint_path} exists on rank {dist.get_global_rank()}? {os.path.exists(latest_checkpoint_path)}',
)
if not latest_checkpoint_exists[0]:
# If the checkpoint doesn't exist on rank 0, don't crash, so the initial autoresume run can succeed
return None
elif not all(latest_checkpoint_exists):
raise RuntimeError('Downloading the checkpoint to all nodes failed')
return latest_checkpoint_path
[docs] def fit(
self,
*,
# Train Dataloader
train_dataloader: Optional[Union[Iterable, DataSpec, Dict[str, Any]]] = None,
train_dataloader_label: str = 'train',
train_subset_num_batches: Optional[int] = None,
spin_dataloaders: Optional[bool] = None,
# Timing
duration: Optional[Union[int, str, Time[int]]] = None,
reset_time: bool = False,
# Schedulers
schedulers: Optional[Union[ComposerScheduler,
LRScheduler,
Sequence[Union[ComposerScheduler, LRScheduler]],
]] = None,
scale_schedule_ratio: float = 1.0,
step_schedulers_every_batch: Optional[bool] = None,
# Evaluation
eval_dataloader: Optional[Union[Iterable, DataSpec, Evaluator, Sequence[Evaluator]]] = None,
eval_subset_num_batches: int = -1,
eval_interval: Union[int, str, Time, Callable[[State, Event], bool]] = 1,
# Numerics
device_train_microbatch_size: Optional[Union[int, float, str]] = None,
precision: Optional[Union[str, Precision]] = None,
):
"""Train the model.
The Composer :class:`.Trainer` supports multiple calls to :meth:`.fit`. Any arguments specified during
the call to :meth:`.fit` will override the values specified when constructing the :class:`.Trainer`.
All arguments are optional, with the following exceptions:
* The ``train_dataloader`` must be specified here if not provided when constructing the :class:`.Trainer`.
* The ``duration`` must be specified here if not provided when constructing the :class:`.Trainer`,
or if this is a subsequent call to :meth:`.fit`.
For example, the following are equivalent:
.. testcode::
# The `train_dataloader` and `duration` can be specified
# when constructing the Trainer
trainer_1 = Trainer(
model=model,
train_dataloader=train_dataloader,
max_duration="1ep",
)
trainer_1.fit()
# Or, these arguments can be specified on `fit()`
trainer_2 = Trainer(model)
trainer_2.fit(
train_dataloader=train_dataloader,
duration="1ep"
)
When invoking :meth:`.fit` for a subsequent time, either ``reset_time`` or ``duration`` must be specified.
Otherwise, it is ambiguous for how long to train.
* If ``reset_time`` is True, then :meth:`.fit` will train for the same amount of time as the previous
call (or for ``duration`` if that parameter is also specified). The :attr:`.State.timestamp` will be reset,
causing :class:`.ComposerScheduler` and :class:`.Algorithm` instances to start from the beginning, as if it
is a new training run. Model gradients, optimizer states, and native PyTorch schedulers will not be reset.
* If ``reset_time`` is False, then :meth:`.fit` will train for the amount of time specified by
``duration``. The :attr:`.State.max_duration` will be incremented by ``duration``.
For example:
.. testcode::
# Construct the trainer
trainer = Trainer(max_duration="1ep")
# Train for 1 epoch
trainer.fit()
assert trainer.state.timestamp.epoch == "1ep"
# Reset the time to 0, then train for 1 epoch
trainer.fit(reset_time=True)
assert trainer.state.timestamp.epoch == "1ep"
# Train for another epoch (2 epochs total)
trainer.fit(duration="1ep")
assert trainer.state.timestamp.epoch == "2ep"
# Train for another batch (2 epochs + 1 batch total)
# It's OK to switch time units!
trainer.fit(duration="1ba")
assert trainer.state.timestamp.epoch == "2ep"
assert trainer.state.timestamp.batch_in_epoch == "1ba"
# Reset the time, then train for 3 epochs
trainer.fit(reset_time=True, duration="3ep")
assert trainer.state.timestamp.epoch == "3ep"
Args:
train_dataloader (Iterable | DataSpec | Dict[str, Any], optional): See :class:`.Trainer`.
train_dataloader_label (str, optional): See :class:`.Trainer`.
train_subset_num_batches (int, optional): See :class:`.Trainer`.
spin_dataloaders (bool, optional): See :class:`.Trainer`.
reset_time (bool): Whether to reset the :attr:`.State.timestamp` to zero values. Defaults to False.
If ``True``, the timestamp will be zeroed out, causing :class:`.ComposerScheduler` and
:class:`.Algorithm` instances to start from the beginning, as if it is a new training run. The model
will be trained for ``duration``, if specified, or for :attr:`.State.max_duration`, which would have
been provided when constructing the :class:`.Trainer` or by a previous call to :meth:`.fit`.
.. note::
Model gradients, optimizer states, and native PyTorch schedulers will not be reset.
If ``False`` (the default), training time will be incremented from where the previous call to
:meth:`.fit` finished (or from zero, if a new training run).
The :attr:`~.State.max_duration` will be incremented by the ``duration`` parameter.
duration (Time[int] | str | int, optional): The 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.
If ``reset_time`` is False (the default), then :attr:`.State.max_duration` will be converted
into the same units as this parameter (if necessary), and then the max duration incremented by the
value of this parameter.
If ``reset_time`` is True, then :attr:`.State.max_duration` will be set to this parameter.
optimizers (torch.optim.Optimizer | Sequence[torch.optim.Optimizer], optional): See :class:`.Trainer`.
schedulers (LRScheduler | ComposerScheduler | Sequence[LRScheduler | ComposerScheduler], optional): See :class:`.Trainer`.
scale_schedule_ratio (float, optional): See :class:`.Trainer`.
step_schedulers_every_batch (bool, optional): See :class:`.Trainer`.
eval_dataloader (Iterable | DataSpec | Evaluator | Sequence[Evaluator], optional): See :class:`.Trainer`.
eval_subset_num_batches (int, optional): See :class:`.Trainer`.
eval_interval (int | str | Time | (State, Event) -> bool, optional): See :class:`.Trainer`.
device_train_microbatch_size (int | float | str, optional): See :class:`.Trainer`.
precision (Precision | str, optional): See :class:`.Trainer`.
"""
# Check Optimizer
if len(self.state.optimizers) == 0:
raise ValueError(
f'No optimizer was specified when constructing the Trainer. As the '
'model had no parameters, SGD was not created by default. This trainer '
'object can only be used to evaluate or predict. Please specify a model '
'with parameters and an optimizer for training.',
)
# Train Dataloader
if train_dataloader is not None:
self._train_data_spec = ensure_data_spec(train_dataloader)
self.state.set_dataloader(self._train_data_spec.dataloader, train_dataloader_label)
self.state.train_dataloader = self.state.dataloader
self.state.device_train_microbatch_size = _get_initial_device_train_microbatch_size(
self.state.device_train_microbatch_size,
self.state.auto_microbatching,
self.state.train_dataloader,
)
if self._train_data_spec is None:
_raise_missing_argument_exception('train_dataloader')
if train_subset_num_batches is not None:
self.state.dataloader_len = train_subset_num_batches
if spin_dataloaders is not None:
self.spin_dataloaders = spin_dataloaders
# Reset Time
if reset_time:
self.state.timestamp = Timestamp()
# Max Duration
if duration is not None:
duration = ensure_time(duration, TimeUnit.EPOCH)
# Effectively increment the max duration (if not resetting the Time)
# or set the max_duration (if resetting the time -- self.state.timestamp.get(duration.unit) will be 0)
# It is important to set the duration, rather than incrementing it, as ``duration`` could be in
# different units than ``max_duration``
self.state.max_duration = duration + self.state.timestamp.get(duration.unit)
if self.state.max_duration is None:
_raise_missing_argument_exception('max_duration')
assert self.state.max_duration is not None
if self.state.dataloader_len is None and self.state.max_duration.unit == TimeUnit.EPOCH:
raise ValueError((
'max_duration cannot be specified in epochs when using an infinite dataloader. Please either '
'provide a dataloader with a length, specify max_duration in batches, samples, or tokens, or provide '
'train_subset_num_batches.'
))
if self.state.max_duration <= self.state.timestamp.get(self.state.max_duration.unit) and not reset_time:
raise ValueError((
f'The max_duration ({self.state.max_duration}) is less than or equal to the elapsed training duration '
f'({self.state.timestamp.get(self.state.max_duration.unit)}). No training would occur. '
'Please provide the `duration` or specify `reset_time=True` in Trainer.fit().'
))
# Scale Schedule Ratio and Schedulers
if scale_schedule_ratio != 1.0:
# Not scaling the schedulers if the ratio is 1.0 in case if the scheduler cannot be scaled
# (e.g. a custom LambdaLR). However, since 1.0 implies no scaling, it is still possible
# to train with it.
self.state.max_duration = _scale_max_duration_by_ssr(scale_schedule_ratio, self.state.max_duration)
if schedulers is not None:
self.state.schedulers = _compile_schedulers(schedulers, self.state, scale_schedule_ratio)
if step_schedulers_every_batch is None:
self._scheduler_step_frequency = _get_default_scheduler_frequency(schedulers)
else:
self._scheduler_step_frequency = TimeUnit.BATCH if step_schedulers_every_batch else TimeUnit.EPOCH
else:
if scale_schedule_ratio != 1.0:
raise ValueError('Specifying `scale_schedule_ratio` without `schedulers` has no effect.')
if step_schedulers_every_batch is not None:
raise ValueError('Specifying `step_schedulers_every_batch` without `schedulers` has no effect.')
if step_schedulers_every_batch is not None:
raise ValueError('Specifying `step_schedulers_every_batch` without `schedulers` has no effect.')
# Evaluators
if eval_dataloader is not None:
# Need to use the `original_model` rather than `state.model`, as `state.model`
# could be DDP / DeepSpeed wrapped.
eval_metrics = self._original_model.get_metrics(is_train=False)
metric_names = [str(k) for k in eval_metrics.keys()]
eval_dataloader = ensure_tuple(eval_dataloader)
evaluator_types = [isinstance(evaluator, Evaluator) for evaluator in eval_dataloader]
if any(evaluator_types) and not all(evaluator_types):
raise ValueError(
'Mixing Evaluator with other classes is not allowed, please wrap'
'all other classes with the Evaluator class. These are the classes'
'that were detected:' + str([type(evaluator) for evaluator in eval_dataloader]),
)
evaluators = [
ensure_evaluator(evaluator, default_metric_names=metric_names) for evaluator in eval_dataloader
]
# match metric names to model metrics
self.state.eval_metrics = {
evaluator.label: _filter_metrics(eval_metrics, evaluator.metric_names) for evaluator in evaluators
}
_set_evaluator_interval_and_subset_num_batches(
evaluators=evaluators,
eval_interval=eval_interval,
subset_num_batches=eval_subset_num_batches,
)
for evaluator in evaluators:
_validate_evaluator(evaluator, self.state.device)
if len(evaluators) == 0:
if eval_subset_num_batches != -1:
warnings.warn(
f'Specifying `eval_subset_num_batches={eval_subset_num_batches}` without an `eval_dataloader` '
'has no effect. If trying to run an evaluator, make sure `eval_dataloader` is specified. '
'Otherwise, set `eval_subset_num_batches` to default value -1.',
)
if eval_interval != 0 and eval_interval != 1:
warnings.warn(
f'Specifying `eval_interval={eval_interval}` without an `eval_dataloader` has no effect. '
'If trying to run an evaluator, make sure `eval_dataloader` is specified. Otherwise, '
'set `eval_interval` to 0 or default value 1.',
)
self.state.evaluators = evaluators
# Microbatching
if device_train_microbatch_size is not None:
self.state.auto_microbatching = _is_auto_microbatching(
device_train_microbatch_size,
device=self.state.device,
)
if self.state.auto_microbatching and self._train_data_spec is not None and hasattr(
self._train_data_spec,
'seq_parallel_world_size',
):
raise ValueError('`device_train_microbatch_size="auto"` is not compatible with sequence parallelism.')
if train_dataloader is not None and hasattr(
train_dataloader,
'seq_parallel_world_size',
) and train_dataloader.seq_parallel_world_size > 1 and abs( # type: ignore
device_train_microbatch_size * train_dataloader.seq_parallel_world_size - 1, # type: ignore
) > 1e-4:
raise ValueError(
'`Sequence parallelism requires a microbatch size of 1 distributed over the sequence parallel group.',
)
if self.state.auto_microbatching and self.state.profiler:
raise ValueError(
"`device_train_microbatch_size='auto'` is not compatible with the profiler. It is "
"recommended to run a mini-run with `device_train_microbatch_size='auto'` to identify "
'the optimal device_train_microbatch_size value and then manually specify that in a '
'second run with profiler.',
)
self.state.device_train_microbatch_size = _get_initial_device_train_microbatch_size(
device_train_microbatch_size,
self.state.auto_microbatching,
self.state.train_dataloader,
)
# Precision
if precision is not None:
if Precision(precision) != self.state.precision:
if self.state.deepspeed_enabled:
raise ValueError('Changing the precision when using DeepSpeed is not supported')
precision = Precision(precision)
_validate_precision(precision, self.state.device)
self.state.precision = precision
# update scaler since precision was provided
self.state.scaler = ClosureGradScaler() if self._use_closures() else GradScaler()
self.first_batch_complete = False
self._train_loop()
[docs] def close(self):
"""Shutdown the trainer.
.. seealso:: :meth:`.Engine.close` for additional information.
"""
self.engine.close()
dist.barrier()
def _ensure_metrics_device_and_dtype(
self,
metrics: Dict[str, Metric],
ensure_cpu: bool = False,
):
for name, metric in metrics.items():
# 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.
# Force all metrics to go on the CPU
if ensure_cpu:
metrics[name] = DeviceCPU().module_to_device(metric)
else:
metrics[name] = self.state.device.module_to_device(metric)
if is_model_deepspeed(self.state.model):
# 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 key in metric._defaults:
metric_data = getattr(metric, key)
if isinstance(metric_data, torch.Tensor) and metric_data.dtype == torch.float16:
metric_data = metric_data.to(torch.float32) # type: ignore
setattr(metric, key, metric_data)
return metrics
def _compute_and_log_metrics(self, dataloader_label: str, metrics: Dict[str, Metric]):
"""Computes metrics, logs the results, and updates the state with the metrics.
Args:
dataloader_label (str): The dataloader label.
metrics (Dict[str, Metric]): The metrics to compute.
"""
# log computed metrics
computed_metrics = {}
for metric_name, metric in metrics.items():
computed_metrics[metric_name] = metric.compute()
self.logger.log_metrics({f'metrics/{dataloader_label}/{name}': val for (name, val) in computed_metrics.items()
},)
# store metric instances
for metric_name, metric in metrics.items():
assert isinstance(metric, Metric)
if dataloader_label == 'train':
assert self.state.train_metrics is not None
self.state.train_metrics[metric_name] = metric
self.state.train_metric_values[metric_name] = computed_metrics[metric_name]
else:
if dataloader_label not in self.state.eval_metrics:
self.state.eval_metrics[dataloader_label] = {}
self.state.eval_metrics[dataloader_label][metric_name] = metric
self.state.eval_metric_values[metric_name] = computed_metrics[metric_name]
def _spin_dataloaders_to_cur_epoch(self):
"""Spin the dataloaders to restore sampler state for current epoch.
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.
"""
log.debug('Spinning the dataloaders')
# Spin the evaluator dataloaders once to initialize its sampler deterministically
# so it does not affect any other RNG reads
eval_state = self.state.dataset_resumption.get('eval', {})
for evaluator in self.state.evaluators:
dataloader = evaluator.dataloader.dataloader
sampler = _get_distributed_sampler(dataloader) if isinstance(dataloader, DataLoader) else None
if isinstance(sampler, DistributedSampler):
sampler.set_epoch(0)
if evaluator.label not in eval_state:
for _ in dataloader:
break
# Spin the train dataloader's sampler to get to the state of the desired epoch
dataloader = self.state.dataloader
assert dataloader is not None, 'train dataloader is set on state after FIT_START'
if 'train' not in self.state.dataset_resumption:
sampler = _get_distributed_sampler(dataloader) if isinstance(dataloader, DataLoader) else None
for epoch in range(int(self.state.timestamp.epoch)):
if isinstance(sampler, DistributedSampler):
sampler.set_epoch(epoch)
for _ in dataloader:
break
def _accumulate_time_across_ranks(
self,
num_samples: Union[int, float],
num_tokens: int,
batch_time: datetime.timedelta,
) -> Tuple[int, int, datetime.timedelta]:
"""Accumulate the number of samples and tokens across ranks.
Returns a (num_samples, num_tokens, batch_time) tuple.
"""
# Samples and tokens should be summed
# Batch time should be the value from rank 0
# num_samples can be floating point if we are doing sequence parallelism, since in that case each rank works on only a part of the sample. For example, with sequence parallelism world size 2, each rank trains on half of a sample.
if isinstance(num_samples, float):
sample_token_tensor = self.state.device.tensor_to_device(
torch.tensor([num_samples, num_tokens], dtype=torch.float32),
)
else:
sample_token_tensor = self.state.device.tensor_to_device(
torch.tensor([num_samples, num_tokens], dtype=torch.int),
)
dist.all_reduce(sample_token_tensor, reduce_operation='SUM')
if isinstance(num_samples, float):
sample_token_tensor_int = sample_token_tensor.round().to(torch.int)
if torch.any(torch.abs(sample_token_tensor_int - sample_token_tensor) > 1e-4):
raise ValueError('The sums of samples and tokens across ranks should each be integers.')
sample_token_tensor = sample_token_tensor_int
batch_time_tensor = self.state.device.tensor_to_device(
torch.tensor([batch_time.total_seconds()], dtype=torch.float32),
)
dist.broadcast(batch_time_tensor, src=0)
batch_time = datetime.timedelta(seconds=batch_time_tensor[0].cpu().item())
return int(sample_token_tensor[0].cpu().item()), int(sample_token_tensor[1].cpu().item()), batch_time
def _train_loop(self) -> None:
"""Run training for the specified number of epochs and log results."""
# Log training start
log.info('Using precision %s', self.state.precision)
self.logger.log_hyperparameters({
'enabled_algorithms/' + algo.__class__.__name__: True for algo in self.state.algorithms
})
assert self.state.dataloader is not None, 'dataloader is set in __init__() or fit()'
assert self._train_data_spec is not None, 'The train data spec is set in __init__() or fit()'
assert self.state.scaler is not None, 'scaler should have been set in __init__()'
self.engine.run_event(Event.FIT_START)
use_grad_scaling = self._use_grad_scaling(self.state.precision, self.state.scaler)
if self.spin_dataloaders:
self._spin_dataloaders_to_cur_epoch()
if self.state.timestamp.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
self.state.model.train()
finished_epoch_early = False
last_wct = datetime.datetime.now()
if self.state.max_duration is None:
# This is essentially just a type check, as max_duration should always be
# asserted to be not None when Trainer.fit() is called
raise RuntimeError('max_duration must be specified when initializing the Trainer')
log.debug('Starting training loop')
while self.state.timestamp < self.state.max_duration:
if int(self.state.timestamp.epoch_in_iteration) == 0 and int(self.state.timestamp.batch_in_epoch) == 0:
self.engine.run_event(Event.ITERATION_START)
if int(self.state.timestamp.batch_in_epoch) == 0:
self.engine.run_event(Event.EPOCH_START)
self.logger.log_metrics({'time/epoch': self.state.timestamp.epoch.value})
dataloader = self.state.dataloader
sampler = _get_distributed_sampler(dataloader) if isinstance(dataloader, DataLoader) else None
if isinstance(sampler, DistributedSampler):
sampler.set_epoch(int(self.state.timestamp.epoch))
for batch_idx, self.state.batch in enumerate(self._iter_dataloader(TrainerMode.TRAIN)):
# Spin dataloader forward unless dataloader handles internally with dataset_resumption
if self.spin_dataloaders and 'train' not in self.state.dataset_resumption and batch_idx < int(
self.state.timestamp.batch_in_epoch,
):
# Restore the RNG state immediately before the next batch is yielded from the dataloader
if batch_idx + 1 == int(self.state.timestamp.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.state.device.batch_to_device(self.state.batch)
self.state.batch = self._train_data_spec.device_transforms(self.state.batch)
rank_num_samples = self._train_data_spec.get_num_samples_in_batch(self.state.batch)
rank_num_tokens = self._train_data_spec.get_num_tokens_in_batch(self.state.batch)
if self.state.deepspeed_enabled:
self.state.batch = _fix_batch_precision_for_deepspeed(self.state.batch, self.state.precision)
self.engine.run_event(Event.AFTER_DATALOADER)
self.engine.run_event(Event.BATCH_START)
# Log time values
self.logger.log_metrics({
'time/batch': self.state.timestamp.batch.value,
'time/sample': self.state.timestamp.sample.value,
'time/batch_in_epoch': self.state.timestamp.batch_in_epoch.value,
'time/sample_in_epoch': self.state.timestamp.sample_in_epoch.value,
})
if rank_num_tokens > 0:
self.logger.log_metrics({'time/token': self.state.timestamp.token.value})
self.logger.log_metrics({'time/token_in_epoch': self.state.timestamp.token_in_epoch.value})
total_loss_dict = self._train_batch(use_grad_scaling)
if use_grad_scaling:
self.state.scaler.update()
# total_loss_dict can be None if gradient scaling failed
if total_loss_dict is not None: # pyright: ignore[reportUnnecessaryComparison]
map_collection(total_loss_dict, dist.all_reduce)
total_loss_dict = {
k: loss.cpu().item() / dist.get_world_size() for k, loss in total_loss_dict.items()
}
self.state.total_loss_dict = total_loss_dict
self.logger.log_metrics(total_loss_dict)
# The scheduler step.step() and compute_and_log_metrics() are going to be included in the
# next batch's wall clock time. The time accumulation must be done here so schedulers
# have the latest timing information
now = datetime.datetime.now()
batch_time = now - last_wct
total_num_samples, total_num_tokens, batch_time = self._accumulate_time_across_ranks(
rank_num_samples,
rank_num_tokens,
batch_time,
)
# `now` is actually in the past, but want to include the time it takes to perform this reduction
last_wct = now
if self._scheduler_step_frequency == TimeUnit.BATCH:
for scheduler in self.state.schedulers:
scheduler.step()
if self.state.train_metrics is not None: # pyright: ignore[reportUnnecessaryComparison]
self._compute_and_log_metrics(
dataloader_label='train',
metrics=self.state.train_metrics,
)
self.state.previous_timestamp = self.state.timestamp
self.state.timestamp = self.state.timestamp.to_next_batch(
samples=total_num_samples,
tokens=total_num_tokens,
duration=batch_time,
)
self.engine.run_event(Event.BATCH_END)
# Pause the timing during evaluation
# Evaluation time is tracked separately in state.eval_timestamp
duration = datetime.datetime.now() - last_wct
self._run_evaluators(Event.BATCH_END)
last_wct = datetime.datetime.now() - duration
self.engine.run_event(Event.BATCH_CHECKPOINT)
if self.state.timestamp >= 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.
finished_epoch_early = True
break
if not finished_epoch_early or self.state.dataloader_len == self.state.timestamp.batch_in_epoch:
# Trigger the epoch end events if the dataloader was exhausted.
# This happens if the "break" did not trigger above, or if it
# did (e.g. duration specified in samples/batches/tokens), but it is still
# the end of the dataloader (i.e. next(dataloader) would raise StopIteration)
if self.state.train_metrics is not None: # pyright: ignore[reportUnnecessaryComparison]
self.state.train_metrics = self._ensure_metrics_device_and_dtype(self.state.train_metrics)
self._compute_and_log_metrics(
dataloader_label='train',
metrics=self.state.train_metrics,
)
if self._scheduler_step_frequency == TimeUnit.EPOCH:
for scheduler in self.state.schedulers:
scheduler.step()
self.state.previous_timestamp = self.state.timestamp
self.state.timestamp = self.state.timestamp.to_next_epoch()
self.engine.run_event(Event.EPOCH_END)
# Pause the timing during evaluation
# Evaluation time is tracked separately in state.eval_timestamp
duration = datetime.datetime.now() - last_wct
self._run_evaluators(Event.EPOCH_END)
last_wct = datetime.datetime.now() - duration
self.engine.run_event(Event.EPOCH_CHECKPOINT)
# Increment iteration
if (
self.state._iteration_length is not None and
self.state.timestamp.epoch_in_iteration == self.state._iteration_length
):
self.state.previous_timestamp = self.state.timestamp
self.state.timestamp = self.state.timestamp.to_next_iteration()
self.engine.run_event(Event.ITERATION_END)
self.engine.run_event(Event.ITERATION_CHECKPOINT)
# Log final time values
self.logger.log_metrics({
'time/epoch': self.state.timestamp.epoch.value,
'time/batch': self.state.timestamp.batch.value,
'time/sample': self.state.timestamp.sample.value,
'time/batch_in_epoch': self.state.timestamp.batch_in_epoch.value,
'time/sample_in_epoch': self.state.timestamp.sample_in_epoch.value,
})
if self.state.timestamp.token.value > 0:
self.logger.log_metrics({'time/token': self.state.timestamp.token.value})
self.logger.log_metrics({'time/token_in_epoch': self.state.timestamp.token_in_epoch.value})
self.engine.run_event(Event.FIT_END)
self._run_evaluators(Event.FIT_END)
def _eval_train_metrics(self, device_batch):
assert self._train_data_spec is not None, 'The train data spec should be set on __init__ or fit()'
assert self.state.train_metrics is not None, 'The train metrics should be set on __init__ or fit()'
with torch.no_grad(),\
model_eval_mode(self.state.model),\
_get_precision_context(self.state.precision, self.state.precision_config, self.state.deepspeed_enabled):
eval_outputs = self._original_model.eval_forward(device_batch, self.state.outputs)
for metric in self.state.train_metrics.values():
self._original_model.update_metric(
device_batch,
eval_outputs,
metric,
)
def _run_evaluators(self, event: Event):
"""Runs evaluators periodically during training."""
evaluators_executing = []
for evaluator in self.state.evaluators:
assert evaluator.eval_interval is not None, 'eval_interval should have been set on __init__() or fit()'
assert evaluator.subset_num_batches is not None, 'subset_num_batches should have been set on __init__() or fit()'
evaluators_executing.append(evaluator.eval_interval(self.state, event))
if not any(evaluators_executing):
return
self.engine.run_event(Event.EVAL_BEFORE_ALL)
for index, evaluator in enumerate(self.state.evaluators):
if evaluators_executing[index]:
self._eval_loop(
evaluator=evaluator,
subset_num_batches=evaluator.subset_num_batches,
metrics=self.state.eval_metrics[evaluator.label],
)
self.engine.run_event(Event.EVAL_AFTER_ALL)
def _train_batch(self, use_grad_scaling: bool) -> Dict[str, torch.Tensor]:
"""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.
Returns:
Dict[str, torch.Tensor]: a dictionary containing the total loss and individual losses if available.
"""
assert self._train_data_spec is not None, 'The train data spec should be set on __init__ or fit()'
# Cache the device batch, because `self.state.batch` gets overridden in microbatching loop.
# Any in-place changes to a microbatch will be reflected in the device batch.
device_batch = self.state.batch
# Retry until we successfully complete training and return loss
while True:
# Reset train_metrics on every batch
# Placing reset here ensures that if auto grad accum catches an OOM, incomplete metric state is cleared
if self.state.train_metrics is not None: # pyright: ignore[reportUnnecessaryComparison]
for metric in self.state.train_metrics.values():
metric.reset()
total_loss_dict = {
'loss/train/total': self.state.device.tensor_to_device(torch.zeros(size=(1,))),
}
found_cuda_oom = 0 # int since bool BOR not supported on all torch.distributed backends
try:
assert self.state.scaler is not None
assert self.state.device_train_microbatch_size is not None
microbatches = self._train_data_spec.split_batch(device_batch, self.state.device_train_microbatch_size)
if self._use_closures():
for optimizer in self.state.optimizers:
if use_grad_scaling:
self.state.scaler.step(
optimizer,
closure=lambda loss_dict=total_loss_dict,
**kwargs: self._train_microbatches(microbatches, loss_dict, **kwargs),
)
else:
optimizer.step(
closure=lambda loss_dict=total_loss_dict,
**kwargs: self._train_microbatches(microbatches, loss_dict, **kwargs).item(),
)
else:
self._train_microbatches(microbatches, total_loss_dict)
if not self.state.deepspeed_enabled:
for optimizer in self.state.optimizers:
if use_grad_scaling:
self.state.scaler.step(optimizer)
else:
optimizer.step()
except RuntimeError as e:
if self.state.auto_microbatching and _is_cuda_oom(e):
log.debug((f"Rank {dist.get_global_rank()} OOM'd."))
found_cuda_oom = 1
elif self.state.auto_microbatching and ('cuda' in str(e).lower() or 'c10' in str(e).lower()):
raise RuntimeError(
textwrap.dedent(
'Encountered non-addressable cuda error while using auto microbatching. '
'If this repeatedly occurs, set `device_train_microbatch_size` manually.',
),
) from e
else:
raise
if self.state.auto_microbatching:
all_ranks_finished = False
while not all_ranks_finished:
# Propagate across all ranks if any rank hit CUDA OOM
found_cuda_oom_tensor = self.state.device.tensor_to_device(
torch.tensor([found_cuda_oom], dtype=torch.uint8),
)
dist.all_reduce(found_cuda_oom_tensor, reduce_operation='MAX')
found_cuda_oom = found_cuda_oom_tensor.item()
# Check if any rank is still not done with the batch. This may happen if only a
# subset of ranks OOM, leaving some batches still in the forward pass
all_ranks_finished_tensor = self.state.device.tensor_to_device(torch.tensor([1], dtype=torch.uint8))
dist.all_reduce(all_ranks_finished_tensor, reduce_operation='MIN')
all_ranks_finished = all_ranks_finished_tensor.item() == 1
if found_cuda_oom == 1:
_adjust_device_train_microbatch_size(self.state)
# Skip return and rerun after handling oom
continue
# Log microbatch and return loss if we've completed without OOMing.
assert self.state.device_train_microbatch_size is not None
self.logger.log_metrics({'trainer/device_train_microbatch_size': self.state.device_train_microbatch_size})
self.first_batch_complete = True
return total_loss_dict
def _train_microbatches(
self,
microbatches: Sequence[Batch],
total_loss_dict: Dict[str, torch.Tensor],
ddp_sync: bool = True,
) -> torch.Tensor:
"""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.
total_loss_dict (Dict[str, torch.tensor]): Dictionary containing individual losses and their sum aggregated across all
microbatches.
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:
if self.state.auto_microbatching and not self.first_batch_complete:
# PyTorch DDP rebuilds gradient reduction buckets after 1) a forward pass where the
# no_sync context was not set 2) a backward pass 3) a forward pass. If only a
# subset of ranks OOM on the first batch, this will cause a deadlock since a rank
# that did not OOM will complete steps (1), (2), and (3) on the first succesful
# microbatch after the OOMs but an OOMing rank will have never completed (1) if
# using `SINGLE_AUTO_SYNC`. To avoid this, we force a sync on every microbatch for
# the first batch.
log.info(
'Auto microbatching requires syncing every microbatch (`MULTI_AUTO_SYNC`)'
' to avoid deadlock on first batch, so ddp_sync_strategy will be ignored.',
)
context = contextlib.nullcontext
else:
context = cast(Callable[[], ContextManager], self.state.model.no_sync)
assert self._train_data_spec is not None
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.state.deepspeed_enabled:
for optimizer in self.state.optimizers:
try:
optimizer.zero_grad(set_to_none=True)
except TypeError:
optimizer.zero_grad()
# Tracker for gradient accumulation
current_batch_size = sum([self._train_data_spec.get_num_samples_in_batch(batch) for batch in microbatches])
# Cache batch, which will be overwritten by microbatches. Restore after microbatches complete
current_batch = self.state.batch
for microbatch_idx, self.state.batch in enumerate(microbatches):
is_final_microbatch = microbatch_idx + 1 == len(microbatches)
microbatch_loss_dict = self._train_microbatch(use_grad_scaling, current_batch_size, is_final_microbatch)
# Aggregate each loss in microbatch_loss_dict into total_loss_dict
for k, microbatch_loss in microbatch_loss_dict.items():
loss_key = f'loss/train/{k}'
if loss_key not in total_loss_dict:
total_loss_dict[loss_key] = self.state.device.tensor_to_device(torch.zeros(size=(1,)))
total_loss_dict[loss_key] += microbatch_loss
# Restore batch
self.state.batch = current_batch
# Unscale gradients before `Event.AFTER_TRAIN_BATCH`
if use_grad_scaling:
for optimizer in ensure_tuple(self.state.optimizers):
self.state.scaler.unscale_(optimizer)
self.engine.run_event(Event.AFTER_TRAIN_BATCH)
return total_loss_dict['loss/train/total']
def _train_microbatch(
self,
use_grad_scaling: bool,
current_batch_size: Union[int, float],
is_final_microbatch: bool,
) -> Dict[str, torch.Tensor]:
"""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.
current_batch_size (int, float): The current batch size.
minibatch_num_samples (int): Number of samples in the minibatch.
is_final_microbatch (bool): If current microbatch is the last one.
"""
assert self.state.scaler is not None
assert self._train_data_spec is not None
# Cache the device batch, because `self.state.batch` gets overridden in microbatching loop
device_batch = deepcopy(self.state.batch)
microbatch_num_samples = self._train_data_spec.get_num_samples_in_batch(self.state.batch)
if self.state.deepspeed_enabled or not isinstance(self.state.model, DistributedDataParallel):
sync_context = contextlib.nullcontext()
elif self.state.auto_microbatching and not self.first_batch_complete:
# PyTorch DDP rebuilds gradient reduction buckets after 1) a forward pass where the
# no_sync context was not set 2) a backward pass 3) a forward pass. If only a
# subset of ranks OOM on the first batch, this will cause a deadlock since a rank
# that did not OOM will complete steps (1), (2), and (3) on the first succesful
# microbatch after the OOMs but an OOMing rank will have never completed (1) if
# using `SINGLE_AUTO_SYNC`. To avoid this, we force a sync on every microbatch for
# the first batch.
log.info(
'Auto microbatching requires syncing every microbatch (`MULTI_AUTO_SYNC`)'
' to avoid deadlock on first batch, so ddp_sync_strategy will be ignored.',
)
sync_context = contextlib.nullcontext()
else:
sync_context = 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 _get_precision_context(
self.state.precision,
self.state.precision_config,
self.state.deepspeed_enabled,
):
self.state.outputs = self.state.model(self.state.batch)
self.engine.run_event(Event.AFTER_FORWARD)
# Check if other ranks OOMed after forward pass when using auto microbatching. This may
# happen when close to memory limit or with uneven memory usage across ranks
if self.state.auto_microbatching:
# Check if any other rank hit an OOM
found_cuda_oom_tensor = self.state.device.tensor_to_device(torch.tensor([0], dtype=torch.uint8))
dist.all_reduce(found_cuda_oom_tensor, reduce_operation='MAX')
found_cuda_oom = found_cuda_oom_tensor.item()
# Signal current rank is still in batch
all_ranks_finished_tensor = self.state.device.tensor_to_device(torch.tensor([0], dtype=torch.uint8))
dist.all_reduce(all_ranks_finished_tensor, reduce_operation='MIN')
if found_cuda_oom == 1:
raise RuntimeError('CUDA out of memory encountered on a different rank')
# Loss
self.engine.run_event(Event.BEFORE_LOSS)
with _get_precision_context(
self.state.precision,
self.state.precision_config,
self.state.deepspeed_enabled,
):
self.state.loss = self._original_model.loss(self.state.outputs, self.state.batch)
assert self.state.loss is not None
self.engine.run_event(Event.AFTER_LOSS)
# Backward Pass
self.engine.run_event(Event.BEFORE_BACKWARD)
microbatch_loss_dict = {}
# If total loss key is present, copy loss
if isinstance(self.state.loss, dict) and ('total' in self.state.loss):
microbatch_loss = self.state.loss['total'] # type: ignore
microbatch_loss_dict = self.state.loss.copy()
# If total loss key is not present, sum individual losses
else:
microbatch_loss = self.state.device.tensor_to_device(torch.zeros(size=(1,)))
for loss in ensure_tuple(self.state.loss):
assert isinstance(loss, torch.Tensor)
microbatch_loss.add_(loss.mean())
# Copy the loss if it is a dictionary
if isinstance(self.state.loss, dict):
microbatch_loss_dict = self.state.loss.copy()
# If not, create a dictionary with generic loss names
elif len(ensure_tuple(self.state.loss)) > 1:
microbatch_loss_dict = {f'loss{i}': loss for i, loss in enumerate(ensure_tuple(self.state.loss))}
# Include total loss
microbatch_loss_dict['total'] = microbatch_loss
# For each loss to log: detach, clone, mean, then multiply by (microbatch size) / (batch size)
for k, loss in microbatch_loss_dict.items():
microbatch_loss_dict[k] = loss.detach().clone().mean() * (microbatch_num_samples / current_batch_size)
if use_grad_scaling:
microbatch_loss = cast(torch.Tensor, self.state.scaler.scale(microbatch_loss)) # type: ignore
if self.state.deepspeed_enabled:
self.state.deepspeed_model.backward(microbatch_loss)
else:
# Scale loss based on the number of samples in the microbatch to maintain gradient numerics
microbatch_loss.mul_(microbatch_num_samples / current_batch_size)
microbatch_loss.backward(create_graph=self._backwards_create_graph)
if self.state.device.dist_backend == 'xla':
# For xla devices, the program between any pair of mark_steps() calls is compiled. With out this, the
# microbatching loop is unrolled, drastically increasing compile time.
xm.mark_step()
self.engine.run_event(Event.AFTER_BACKWARD)
# Use microbatch outputs to update training metrics
if (
self.state.train_metrics is not None and # pyright: ignore[reportUnnecessaryComparison]
len(self.state.train_metrics) != 0
):
self.state.train_metrics = self._ensure_metrics_device_and_dtype(self.state.train_metrics)
self._eval_train_metrics(device_batch)
if self.state.deepspeed_enabled:
self.state.deepspeed_model.step()
return microbatch_loss_dict
[docs] def predict(
self,
dataloader: Union[DataLoader, DataSpec],
subset_num_batches: int = -1,
*,
return_outputs: bool = True,
):
"""Output model prediction on the provided data.
There are two ways to access the prediction outputs.
1. With ``return_outputs`` set to True, the batch predictions will be collected into a list and returned.
2. Via a custom callback, which can be used with ``return_outputs`` set to False.
This technique can be useful if collecting all the outputs from the dataloader would exceed available memory,
and you want to write outputs directly to files. For example:
.. testsetup::
predict_dl = train_dataloader
.. testcode::
import os
import torch
from torch.utils.data import DataLoader
from composer import Trainer, Callback
from composer.loggers import Logger
class PredictionSaver(Callback):
def __init__(self, folder: str):
self.folder = folder
os.makedirs(self.folder, exist_ok=True)
def predict_batch_end(self, state: State, logger: Logger) -> None:
name = f'batch_{int(state.predict_timestamp.batch)}.pt'
filepath = os.path.join(self.folder, name)
torch.save(state.outputs, filepath)
# Also upload the files
logger.upload_file(remote_file_name=name, file_path=filepath)
trainer = Trainer(
...,
callbacks=PredictionSaver('./predict_outputs'),
)
trainer.predict(predict_dl, return_outputs=False)
print(sorted(os.listdir('./predict_outputs')))
.. testoutput::
['batch_1.pt', ...]
Args:
dataloader (DataLoader | DataSpec): The :class:`.DataLoader` or
:class:`.DataSpec` for the prediction data.
subset_num_batches (int, optional): If specified, only perform model prediction
on this many batches. This parameter has no effect if it is greater than ``len(dataloader)``.
If ``-1``, then the entire loader will be iterated over. (default: ``-1``)
return_outputs (bool, optional): If True (the default), then prediction outputs will be (recursively)
moved to cpu and accumulated into a list. Otherwise, prediction outputs are discarded after each
batch.
Returns:
List: A list of batch outputs, if ``return_outputs`` is True. Otherwise, an empty list.
"""
if isinstance(dataloader, DataSpec):
data_spec = dataloader
else:
data_spec = DataSpec(dataloader)
# Bind the dataloader to the state, but be able to restore the previous dataloader afterwards
original_dataloader = self.state.dataloader
original_dataloader_label = self.state.dataloader_label
original_dataloader_len = self.state.dataloader_len
self.state.set_dataloader(data_spec.dataloader, 'predict', subset_num_batches)
assert self.state.dataloader is not None, 'Already set the dataloader'
# Reset the predict timestamp
self.state.predict_timestamp = Timestamp()
last_wct = datetime.datetime.now()
outputs = []
cpu_device = DeviceCPU()
with torch.no_grad(), model_eval_mode(self.state.model):
self.engine.run_event(Event.PREDICT_START)
for self.state.batch in self._iter_dataloader(TrainerMode.PREDICT):
# Move the batch onto the device
self.state.batch = self.state.device.batch_to_device(self.state.batch)
# Perform any device transforms
self.state.batch = data_spec.device_transforms(self.state.batch)
# Count the batch size and num tokens before any events run
rank_num_samples = data_spec.get_num_samples_in_batch(self.state.batch)
rank_num_tokens = data_spec.get_num_tokens_in_batch(self.state.batch)
# Fix the batch if using DeepSpeed
if self.state.deepspeed_enabled:
self.state.batch = _fix_batch_precision_for_deepspeed(self.state.batch, self.state.precision)
self.engine.run_event(Event.PREDICT_BATCH_START)
self.engine.run_event(Event.PREDICT_BEFORE_FORWARD)
with _get_precision_context(
self.state.precision,
self.state.precision_config,
self.state.deepspeed_enabled,
):
self.state.outputs = self.state.model(self.state.batch)
self.engine.run_event(Event.PREDICT_AFTER_FORWARD)
if return_outputs:
outputs.append(cpu_device.batch_to_device(self.state.outputs))
now = datetime.datetime.now()
batch_time = now - last_wct
total_num_samples, total_num_tokens, batch_time = self._accumulate_time_across_ranks(
num_samples=rank_num_samples,
num_tokens=rank_num_tokens,
batch_time=batch_time,
)
last_wct = now
self.state.predict_timestamp = self.state.predict_timestamp.to_next_batch(
samples=total_num_samples,
tokens=total_num_tokens,
duration=batch_time,
)
self.engine.run_event(Event.PREDICT_BATCH_END)
self.engine.run_event(Event.PREDICT_END)
# Restore the dataloader
self.state.set_dataloader(original_dataloader, original_dataloader_label)
if original_dataloader_len is not None:
self.state.dataloader_len = original_dataloader_len
return outputs
[docs] def eval(
self,
eval_dataloader: Optional[Union[Iterable, DataSpec, Evaluator, Sequence[Evaluator]]] = None,
subset_num_batches: int = -1,
):
"""Run evaluation loop.
Results are stored in ``trainer.state.eval_metrics``. The ``eval_dataloader`` can be provided to
either the eval() method or during training init().
Examples:
.. testcode::
trainer = Trainer(
model=model,
train_dataloader=train_dataloader,
max_duration="2ep",
device="cpu",
)
trainer.fit()
# run eval
trainer.eval(
eval_dataloader=eval_dataloader,
)
Or, if the ``eval_dataloader`` is provided during init:
.. testcode::
trainer = Trainer(
model=model,
eval_dataloader=eval_dataloader,
train_dataloader=train_dataloader,
max_duration="2ep",
device="cpu",
)
trainer.fit()
# eval_dataloader already provided:
trainer.eval()
For multiple metrics or dataloaders, use :class:`.Evaluator` to provide
identifier names. For example, to run the GLUE task:
.. code:: python
from composer.core import Evaluator
from composer.models.nlp_metrics import BinaryF1Score
glue_mrpc_task = Evaluator(
label='glue_mrpc',
dataloader=mrpc_dataloader,
metric_names=['BinaryF1Score', 'MulticlassAccuracy']
)
glue_mnli_task = Evaluator(
label='glue_mnli',
dataloader=mnli_dataloader,
metric_names=['MulticlassAccuracy']
)
trainer = Trainer(
...,
eval_dataloader=[glue_mrpc_task, glue_mnli_task],
...
)
The metrics used are defined in your model's ``get_metrics()`` method. For more information,
see :doc:`/trainer/evaluation`.
.. note::
If evaluating with multiple GPUs using a DistributedSampler with `drop_last=False`, the last
batch will contain duplicate samples, which may affect metrics. To avoid this, as long as
the dataset passed to the DistributedSampler has a length defined, Composer will correctly
drop duplicate samples.
Args:
eval_dataloader (Iterable | DataLoader | DataSpec | Evaluator | Sequence[Evaluator], optional): Dataloaders
for evaluation. If not provided, defaults to using the
``eval_dataloader`` provided to the trainer init().
subset_num_batches (int, optional): Evaluate on this many batches. Default to ``-1`` (the entire
dataloader. Can also be provided in the trainer.__init__() as ``eval_subset_num_batches``.
"""
self.engine.run_event(Event.EVAL_STANDALONE_START)
if eval_dataloader is not None:
eval_passed_in = True
eval_metrics = deepcopy(self._original_model.get_metrics(is_train=False))
metric_names = [str(k) for k in eval_metrics.keys()]
eval_dataloader = ensure_tuple(eval_dataloader)
evaluator_types = [isinstance(evaluator, Evaluator) for evaluator in eval_dataloader]
if any(evaluator_types) and not all(evaluator_types):
raise ValueError(
'Mixing Evaluator with other classes is not allowed, please wrap'
'all other classes with the Evaluator class. These are the classes'
'that were detected:' + str([type(evaluator) for evaluator in eval_dataloader]),
)
evaluators = [
ensure_evaluator(evaluator, default_metric_names=metric_names) for evaluator in eval_dataloader
]
if self.state.eval_metrics:
for evaluator in evaluators:
if evaluator.label in self.state.eval_metrics:
warnings.warn(
f'eval_dataloader label \'{evaluator.label}\' was already provided in '
'trainer initialization. Existing data for that label will be overwritten. '
'To prevent this in the future, assign unique label names.',
category=UserWarning,
)
# match metric names to model metrics
log.info(f'Added {[e.label for e in evaluators]} to eval_metrics.')
self.state.eval_metrics.update({e.label: _filter_metrics(eval_metrics, e.metric_names) for e in evaluators})
_set_evaluator_interval_and_subset_num_batches(
evaluators=evaluators,
eval_interval='1ep', # ignored
subset_num_batches=subset_num_batches,
)
for evaluator in evaluators:
_validate_evaluator(evaluator, self.state.device)
self.state.evaluators.extend(evaluators) # Add evaluators to state.evaluators
else:
eval_passed_in = False
if not self.state.evaluators:
raise ValueError('eval_dataloader must be provided to either Trainer init() or eval().')
evaluators = self.state.evaluators
for evaluator in evaluators:
eval_subset_num_batches = evaluator.subset_num_batches if subset_num_batches == -1 else subset_num_batches
self._eval_loop(
evaluator=evaluator,
metrics=self.state.eval_metrics[evaluator.label],
subset_num_batches=eval_subset_num_batches,
)
if eval_passed_in:
self.state.evaluators.remove(evaluator) # Remove them from state once eval is finished.
self.engine.run_event(Event.EVAL_STANDALONE_END)
def _eval_loop(
self,
evaluator: Evaluator,
metrics: Dict[str, Metric],
subset_num_batches: Optional[int] = None,
):
"""Evaluate the model and log appropriate metrics.
Args:
evaluator (Evaluator): The evaluator to use for evaluation.
metrics (Dict[str, Metric]): Dictionary mapping metric names to metrics to evaluate against.
subset_num_batches (int, optional): If specified, evaluate on this many batches. Defaults to ``-1``,
which means to iterate over the entire dataloader.
"""
if subset_num_batches is None:
subset_num_batches = -1
# back up the original dataloader on the state, so we can restore it after evaluation is finished
original_dataloader = self.state.dataloader
original_dataloader_label = self.state.dataloader_label
original_num_batches = self.state.dataloader_len
# Unpack data_spec
data_spec = evaluator.dataloader
# Reset the eval timestamp
self.state.eval_timestamp = Timestamp()
last_wct = datetime.datetime.now()
with torch.no_grad(), model_eval_mode(self.state.model):
self.state.set_dataloader(data_spec.dataloader, evaluator.label, subset_num_batches)
assert self.state.dataloader is not None, 'dataloader is set'
self.engine.run_event(Event.EVAL_START)
# On MPS device we ensure the eval metrics are computed on CPU to avoid numerical errors
metrics = self._ensure_metrics_device_and_dtype(
metrics,
ensure_cpu=isinstance(self.state.device, DeviceMPS),
)
for metric in metrics.values():
metric.reset()
dataloader = self.state.dataloader
drop_last = None
dataset_len = None
last_batch = False
dist_sampler = _get_distributed_sampler(dataloader) if isinstance(dataloader, DataLoader) else None
if isinstance(dist_sampler, DistributedSampler) and isinstance(dataloader, DataLoader):
# 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.
dist_sampler.set_epoch(int(self.state.timestamp.batch))
drop_last = dataloader.drop_last
# Only compute the dataset length if drop_last is False, as otherwise we don't need
# to remove any duplicate samples.
if drop_last == False:
try:
dataset_len = len(dist_sampler.dataset) # type: ignore
except AttributeError:
warnings.warn(
"DistributedSampler's dataset does not have length defined. When "
'`drop_last=False`, metrics may be incorrect, as DistributedSampler '
'duplicates samples to make the dataset divisible by world size. To '
'fix this, provide a dataset with a length attribute to the '
'DistributedSampler to correctly drop duplicate samples.',
)
for self.state.batch in self._iter_dataloader(TrainerMode.EVAL):
self.state.batch = self.state.device.batch_to_device(self.state.batch)
self.state.batch = data_spec.device_transforms(self.state.batch)
# Count the batch size and num tokens before any events run
rank_num_samples = data_spec.get_num_samples_in_batch(self.state.batch)
rank_num_tokens = data_spec.get_num_tokens_in_batch(self.state.batch)
# If using a distributed sampler, keep track of last_batch for metrics update
if dist_sampler is not None and drop_last == False and dataset_len is not None:
batch_num_samples_tensor = self.state.device.tensor_to_device(torch.tensor(rank_num_samples))
dist.all_reduce(batch_num_samples_tensor, reduce_operation='SUM')
batch_num_samples = int(batch_num_samples_tensor.item())
if abs(batch_num_samples - batch_num_samples_tensor.item()) > 1e-4:
raise ValueError('Number of samples in a batch should be an integer.')
last_batch = self.state.eval_timestamp.sample + batch_num_samples >= dataset_len
if self.state.deepspeed_enabled:
self.state.batch = _fix_batch_precision_for_deepspeed(self.state.batch, self.state.precision)
self.engine.run_event(Event.EVAL_BATCH_START)
# Cache the device batch, because `self.state.batch` gets overridden in microbatching loop
device_batch = self.state.batch
# Retry until we successfully complete evaluation
while True:
# Note: We use uint8 instead of bool as BOR is not supported on all torch.distributed backends
found_cuda_oom = 0
try:
microbatches = data_spec.split_batch(device_batch, evaluator.device_eval_microbatch_size)
for i, self.state.batch in enumerate(microbatches):
last_microbatch = i == len(microbatches) - 1
skip_metric_update = False
# Distributed samplers pad batches to be the same size. If using a
# distributed sampler and on last batch, remove the padding
if dist_sampler is not None and drop_last == False and dataset_len is not None and last_batch and last_microbatch:
padding = dist_sampler.total_size - dataset_len
if dist.get_global_rank() >= dist.get_world_size() - padding:
rank_num_samples -= 1
num_samples_in_microbatch = data_spec.get_num_samples_in_batch(self.state.batch)
# Skip updating metric if batch is only padded samples
if num_samples_in_microbatch == 1 or hasattr(data_spec, 'seq_parallel_world_size'):
skip_metric_update = True
# Remove padded samples from batch
else:
if not isinstance(num_samples_in_microbatch, int):
raise ValueError('Number of samples in a batch should be an integer.')
self.state.batch = data_spec.split_batch(
self.state.batch,
num_samples_in_microbatch - 1,
)[0]
self.engine.run_event(Event.EVAL_BEFORE_FORWARD)
with _get_precision_context(
self.state.precision,
self.state.precision_config,
self.state.deepspeed_enabled,
):
self.state.outputs = self._original_model.eval_forward(self.state.batch)
self.engine.run_event(Event.EVAL_AFTER_FORWARD)
# Skip metric update if batch is only padded samples. We do this after
# forward as all models must run forward for FSDP.
if skip_metric_update:
continue
# Run in same precision context to avoid NaNs
with _get_precision_context(
self.state.precision,
self.state.precision_config,
self.state.deepspeed_enabled,
):
if isinstance(self.state.device, DeviceMPS):
# torchmetrics math has numerical errors on M1 devices
# running the compute on CPU instead
if isinstance(self.state.outputs, Mapping):
outputs = {}
for k, v in self.state.outputs.items():
if isinstance(v, torch.Tensor):
outputs[k] = v.cpu()
else:
outputs[k] = v
elif isinstance(self.state.outputs, Sequence):
outputs = []
for v in self.state.outputs:
if isinstance(v, torch.Tensor):
outputs.append(v.cpu())
else:
outputs.append(v)
else:
outputs = self.state.outputs.cpu()
batch = DeviceCPU().batch_to_device(self.state.batch,)
else:
outputs = self.state.outputs
batch = self.state.batch
for metric in metrics.values():
metric_outputs = self._original_model.update_metric(
batch,
outputs,
metric,
)
self.state.metric_outputs = metric_outputs or {}
except RuntimeError as e:
if evaluator.auto_microbatching and _is_cuda_oom(e):
log.debug((f"Rank {dist.get_global_rank()} OOM'd."))
found_cuda_oom = 1
elif self.state.auto_microbatching and ('cuda' in str(e).lower() or 'c10' in str(e).lower()):
raise ValueError(
textwrap.dedent(
'Encountered non-addressable cuda error while using auto microbatching. '
'If this repeatedly occurs, set `device_eval_microbatch_size` manually.',
),
) from e
else:
raise
if evaluator.auto_microbatching:
# Propagate across all ranks if any rank hit CUDA OOM
found_cuda_oom = self.state.device.tensor_to_device(
torch.tensor([found_cuda_oom], dtype=torch.uint8),
)
dist.all_reduce(found_cuda_oom, reduce_operation='MAX')
if found_cuda_oom.item() == 1:
_adjust_device_eval_microbatch_size(evaluator)
# Skip return and rerun after handling oom
continue
# Log device_eval_microbatch_size if auto_microbatching is enabled
self.logger.log_metrics({
f'trainer/{evaluator.label}/device_eval_microbatch_size':
evaluator.device_eval_microbatch_size,
})
# Break if we've successfully completed eval without OOMing.
break
now = datetime.datetime.now()
batch_time = now - last_wct
total_num_samples, total_num_tokens, batch_time = self._accumulate_time_across_ranks(
num_samples=rank_num_samples,
num_tokens=rank_num_tokens,
batch_time=batch_time,
)
self.state.eval_timestamp = self.state.eval_timestamp.to_next_batch(
samples=total_num_samples,
tokens=total_num_tokens,
duration=batch_time,
)
last_wct = now
self.engine.run_event(Event.EVAL_BATCH_END)
self._compute_and_log_metrics(dataloader_label=evaluator.label, metrics=metrics)
self.engine.run_event(Event.EVAL_END)
self.state.set_dataloader(original_dataloader, original_dataloader_label)
if original_num_batches is not None:
self.state.dataloader_len = original_num_batches
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.state.deepspeed_enabled:
return False
precision = Precision(precision)
use_grad_scaling = precision == Precision.AMP_FP16
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 _iter_dataloader(self, trainer_mode: TrainerMode):
"""Helper method to iterate over the dataloader.
This method yields up to :attr:`.State.dataloader_len`` batches from the dataloader. In addition, if the
profiler is enabled, the dataloader latency recorded via the :class:`.Marker` API.
Args:
trainer_mode (TrainerMode): Specifies which mode the trainer is in.
"""
assert self.state.dataloader is not None, 'the dataloader should be set before calling this method'
if self.state.dataloader_len is None:
dataloader_iter = iter(self.state.dataloader)
else:
dataloader_iter = itertools.islice(self.state.dataloader, int(self.state.dataloader_len))
while True:
try:
# [BEFORE/AFTER]_DATALOADER only runs while training
if trainer_mode == TrainerMode.TRAIN:
self.engine.run_event(Event.BEFORE_DATALOADER)
batch = next(dataloader_iter)
except StopIteration:
# [BEFORE/AFTER]_DATALOADER only runs while training
if trainer_mode == TrainerMode.TRAIN:
# Event.AFTER_DATALOADER is normally called in the train loop. However, if we
# encounter StopIteration, the train loop will not run. Accordingly, we need to
# explicitly call the engine to run marker.finish() for the dataloader marker.
# Otherwise, we will encounter an error at the start of the next epoch when
# Event.BEFORE_DATALOADER tries to start an unfinished marker.
self.engine.run_marker_only_event(Event.AFTER_DATALOADER)
break
yield batch
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.state.deepspeed_enabled:
return False
if self.state.device.dist_backend == 'xla':
return False
if self.state.precision != Precision.AMP_FP16:
return True
if not hasattr(self.state, 'optimizers'):
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:
str or None: See :func:`.save_checkpoint`.
"""
return checkpoint.save_checkpoint(
state=self.state,
filename=name,
weights_only=weights_only,
)
[docs] def save_checkpoint_to_save_folder(self):
"""Checkpoints the training :class:`~.State` using a CheckpointSaver if it exists.
Raises:
ValueError: If ``_checkpoint_saver`` does not exist.
Returns:
None
"""
if self._checkpoint_saver is None:
raise ValueError(
'In order to use save_checkpoint_to_save_folder you must pass a save_folder to the Trainer.',
)
else:
self._checkpoint_saver._save_checkpoint(self.state, self.logger)
[docs] def export_for_inference(
self,
save_format: Union[str, ExportFormat],
save_path: str,
save_object_store: Optional[ObjectStore] = None,
sample_input: Optional[Any] = None,
transforms: Optional[Sequence[Transform]] = None,
input_names: Optional[Sequence[str]] = None,
output_names: Optional[Sequence[str]] = None,
):
"""Export a model for inference.
Args:
save_format (Union[str, ExportFormat]): Format to export to. Either ``"torchscript"`` or ``"onnx"``.
save_path: (str): The path for storing the exported model. It can be a path to a file on the local disk,
a URL, or if ``save_object_store`` is set, the object name
in a cloud bucket. For example, ``my_run/exported_model``.
save_object_store (ObjectStore, optional): If the ``save_path`` is in an object name in a cloud bucket
(i.e. AWS S3 or Google Cloud Storage), an instance of
:class:`~.ObjectStore` which will be used
to store the exported model. If this is set to ``None``, will save to ``save_path`` using the trainer's
logger. (default: ``None``)
sample_input (Any, optional): Example model inputs used for tracing. This is needed for "onnx" export.
The ``sample_input`` need not match the batch size you intend to use for inference. However, the model
should accept the ``sample_input`` as is. (default: ``None``)
transforms (Sequence[Transform], optional): transformations (usually optimizations) that should
be applied to the model. Each Transform should be a callable that takes a model and returns a modified model.
input_names (Sequence[str], optional): names to assign to the input nodes of the graph, in order. If set
to ``None``, the keys from the `sample_input` will be used. Fallbacks to ``["input"]``.
output_names (Sequence[str], optional): names to assign to the output nodes of the graph, in order. It set
to ``None``, it defaults to ``["output"]``.
Returns:
None
"""
export_model = self.state.model.module if self.state.is_model_ddp else self.state.model
if not isinstance(export_model, nn.Module):
raise ValueError(f'Exporting Model requires type torch.nn.Module, got {type(export_model)}')
if sample_input == None and save_format == 'onnx':
sample_input = self.state.batch
export_with_logger(
model=export_model,
save_format=save_format,
save_path=save_path,
logger=self.logger,
save_object_store=save_object_store,
sample_input=(sample_input, {}),
transforms=transforms,
input_names=input_names,
output_names=output_names,
)