# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from abc import ABC, abstractmethod
from contextlib import ExitStack
from typing import Any, Callable, ContextManager, Dict, Iterable, List, Optional, Tuple, TypeVar, Union
import torch
from torch import Tensor
from torch.nn import Module
from torch.optim import Optimizer
from torch.utils.data import DataLoader
from lightning.fabric.accelerators import Accelerator
from lightning.fabric.plugins.io.checkpoint_io import CheckpointIO
from lightning.fabric.plugins.io.torch_io import TorchCheckpointIO
from lightning.fabric.plugins.precision import Precision
from lightning.fabric.strategies.launchers.launcher import _Launcher
from lightning.fabric.strategies.registry import _StrategyRegistry
from lightning.fabric.utilities.apply_func import move_data_to_device
from lightning.fabric.utilities.imports import _TORCH_GREATER_EQUAL_2_0
from lightning.fabric.utilities.init import _EmptyInit
from lightning.fabric.utilities.types import _PATH, Optimizable, ReduceOp, _Stateful
TBroadcast = TypeVar("TBroadcast")
TReduce = TypeVar("TReduce")
log = logging.getLogger(__name__)
[docs]class Strategy(ABC):
"""Base class for all strategies that change the behaviour of the training, validation and test- loop."""
def __init__(
self,
accelerator: Optional[Accelerator] = None,
checkpoint_io: Optional[CheckpointIO] = None,
precision: Optional[Precision] = None,
) -> None:
self._accelerator: Optional[Accelerator] = accelerator
self._checkpoint_io: Optional[CheckpointIO] = checkpoint_io
self._precision: Optional[Precision] = None
# Call the precision setter for input validation
self.precision = precision # type: ignore[assignment]
self._launcher: Optional[_Launcher] = None
self._backward_sync_control: Optional[_BackwardSyncControl] = None
@property
@abstractmethod
def root_device(self) -> torch.device:
"""Returns the root device."""
@property
@abstractmethod
def is_global_zero(self) -> bool:
"""Whether the current process is the rank zero process not only on the local node, but for all nodes."""
@property
def launcher(self) -> Optional[_Launcher]:
return self._launcher
@property
def accelerator(self) -> Optional[Accelerator]:
return self._accelerator
@accelerator.setter
def accelerator(self, accelerator: Accelerator) -> None:
self._accelerator = accelerator
@property
def checkpoint_io(self) -> CheckpointIO:
if self._checkpoint_io is None:
self._checkpoint_io = TorchCheckpointIO()
return self._checkpoint_io
@checkpoint_io.setter
def checkpoint_io(self, io: CheckpointIO) -> None:
self._checkpoint_io = io
@property
def precision(self) -> Precision:
return self._precision if self._precision is not None else Precision()
@precision.setter
def precision(self, precision: Optional[Precision]) -> None:
self._precision = precision
[docs] def setup_environment(self) -> None:
"""Setup any processes or distributed connections.
This must be called by the framework at the beginning of every process, before any distributed communication
takes place.
"""
assert self.accelerator is not None
self.accelerator.setup_device(self.root_device)
[docs] def process_dataloader(self, dataloader: DataLoader) -> DataLoader:
"""Wraps the dataloader if necessary.
Args:
dataloader: iterable. Ideally of type: :class:`torch.utils.data.DataLoader`
"""
return dataloader
[docs] def tensor_init_context(self) -> ContextManager:
"""Controls how tensors get created (device, dtype)."""
precision_init_ctx = self.precision.tensor_init_context()
stack = ExitStack()
if _TORCH_GREATER_EQUAL_2_0:
stack.enter_context(self.root_device)
stack.enter_context(precision_init_ctx)
return stack
[docs] def module_init_context(self, empty_init: Optional[bool] = None) -> ContextManager:
"""A context manager wrapping the model instantiation.
Here, the strategy can control how the parameters of the model get created (device, dtype) and or apply other
patches to the model.
Args:
empty_init: Whether to initialize the model with empty weights (uninitialized memory).
If ``None``, the strategy will decide. Some strategies may not support all options.
"""
precision_module_ctx = self.precision.module_init_context()
stack = ExitStack()
if _TORCH_GREATER_EQUAL_2_0:
stack.enter_context(self.root_device)
stack.enter_context(_EmptyInit(enabled=bool(empty_init)))
stack.enter_context(precision_module_ctx)
return stack
[docs] def setup_module_and_optimizers(
self, module: Module, optimizers: List[Optimizer]
) -> Tuple[Module, List[Optimizer]]:
"""Set up a model and multiple optimizers together.
The returned objects are expected to be in the same order they were passed in. The default implementation will
call :meth:`setup_module` and :meth:`setup_optimizer` on the inputs.
"""
module = self.setup_module(module)
optimizers = [self.setup_optimizer(optimizer) for optimizer in optimizers]
return module, optimizers
[docs] def setup_module(self, module: Module) -> Module:
"""Performs setup for the model, e.g., by wrapping it by another class."""
return module
[docs] def setup_optimizer(self, optimizer: Optimizer) -> Optimizer:
"""Performs setup for the optimizer, e.g., by wrapping it by another class."""
return optimizer
[docs] @abstractmethod
def module_to_device(self, module: Module) -> None:
"""Moves the model to the correct device."""
[docs] def batch_to_device(self, batch: Any, device: Optional[torch.device] = None) -> Any:
"""Moves the batch to the correct device.
The returned batch is of the same type as the input batch, just
having all tensors on the correct device.
Args:
batch: The batch of samples to move to the correct device
device: The target device
"""
device = device or self.root_device
return move_data_to_device(batch, device)
[docs] def backward(self, tensor: Tensor, module: Optional[Module], *args: Any, **kwargs: Any) -> None:
r"""Forwards backward-calls to the precision plugin."""
self.precision.pre_backward(tensor, module)
self.precision.backward(tensor, module, *args, **kwargs)
self.precision.post_backward(tensor, module)
[docs] def optimizer_step(
self,
optimizer: Optimizable,
**kwargs: Any,
) -> Any:
"""Performs the actual optimizer step.
Args:
optimizer: the optimizer performing the step
**kwargs: Any extra arguments to ``optimizer.step``
"""
return self.precision.optimizer_step(optimizer, **kwargs)
[docs] @abstractmethod
def all_gather(self, tensor: Tensor, group: Optional[Any] = None, sync_grads: bool = False) -> Tensor:
"""Perform an all_gather on all processes.
Args:
tensor: the tensor to all_gather
group: the process group to gather results from
sync_grads: flag that allows users to synchronize gradients for all_gather op
"""
[docs] @abstractmethod
def all_reduce(
self,
tensor: Union[Tensor, Any],
group: Optional[Any] = None,
reduce_op: Optional[Union[ReduceOp, str]] = "mean",
) -> Union[Tensor, Any]:
"""Reduces the given tensor (e.g. across GPUs/processes).
Args:
tensor: the tensor to sync and reduce
group: the process group to reduce
reduce_op: the reduction operation. Defaults to 'mean'.
Can also be a string 'sum' or ReduceOp.
"""
[docs] @abstractmethod
def barrier(self, name: Optional[str] = None) -> None:
"""Synchronizes all processes which blocks processes until the whole group enters this function.
Args:
name: an optional name to pass into barrier.
"""
[docs] @abstractmethod
def broadcast(self, obj: TBroadcast, src: int = 0) -> TBroadcast:
"""Broadcasts an object to all processes.
Args:
obj: the object to broadcast
src: source rank
"""
[docs] def reduce_boolean_decision(self, decision: bool, all: bool = True) -> bool:
"""Reduce a boolean decision across all processes."""
return decision
[docs] def save_checkpoint(
self,
path: _PATH,
state: Dict[str, Union[Module, Optimizer, Any]],
storage_options: Optional[Any] = None,
filter: Optional[Dict[str, Callable[[str, Any], bool]]] = None,
) -> None:
"""Save model, optimizer, and other state as a checkpoint file.
Args:
path: A path to where the file(s) should be saved
state: A dictionary with contents to be saved. If the dict contains modules or optimizers, their
state-dict will be retrieved and converted automatically.
storage_options: Additional options for the ``CheckpointIO`` plugin
filter: An optional dictionary containing filter callables that return a boolean indicating whether the
given item should be saved (``True``) or filtered out (``False``). Each filter key should match a
state key, where its filter will be applied to the ``state_dict`` generated.
"""
state = self._convert_stateful_objects_in_state(state, filter=(filter or {}))
if self.is_global_zero:
self.checkpoint_io.save_checkpoint(checkpoint=state, path=path, storage_options=storage_options)
[docs] def get_module_state_dict(self, module: Module) -> Dict[str, Union[Any, Tensor]]:
"""Returns model state."""
return module.state_dict()
[docs] def load_module_state_dict(
self, module: Module, state_dict: Dict[str, Union[Any, Tensor]], strict: bool = True
) -> None:
"""Loads the given state into the model."""
module.load_state_dict(state_dict, strict=strict)
[docs] def get_optimizer_state(self, optimizer: Optimizer) -> Dict[str, Tensor]:
"""Returns state of an optimizer.
Allows for syncing/collating optimizer state from processes in custom plugins.
"""
if hasattr(optimizer, "consolidate_state_dict"):
# there are optimizers like PyTorch's ZeroRedundancyOptimizer that shard their
# states, and to avoid OOM we consolidate the full state on rank 0 only
optimizer.consolidate_state_dict()
return optimizer.state_dict() if self.is_global_zero else {}
# for optimizers that are not sharded, we return the state dict on all ranks
return optimizer.state_dict()
[docs] def load_checkpoint(
self,
path: _PATH,
state: Optional[Union[Module, Optimizer, Dict[str, Union[Module, Optimizer, Any]]]] = None,
strict: bool = True,
) -> Dict[str, Any]:
"""Load the contents from a checkpoint and restore the state of the given objects.
Args:
path: A path to where the file is located
state: Can be one of:
- A dictionary of objects whose state will be restored in-place from the checkpoint path.
- ``None`` or the empty dict: The loaded checkpoint will be returned in full.
- A :class:`~torch.nn.Module` instance, if the checkpoint file contains a raw module state dict.
- A :class:`~torch.optim.Optimizer` instance, if the checkpoint file contains a raw optimizer state.
strict: Whether to enforce that the keys in `state` match the keys in the checkpoint.
Returns:
The remaining items that were not restored into the given state dictionary. If no state dictionary is
given, the full checkpoint will be returned.
"""
torch.cuda.empty_cache()
checkpoint = self.checkpoint_io.load_checkpoint(path)
if not state:
return checkpoint
if isinstance(state, Module):
self.load_module_state_dict(module=state, state_dict=checkpoint, strict=strict)
return {}
if isinstance(state, Optimizer):
state.load_state_dict(checkpoint)
return {}
_validate_keys_for_strict_loading(state.keys(), checkpoint.keys(), strict=strict)
for name, obj in state.copy().items():
if name not in checkpoint:
continue
if isinstance(obj, _Stateful):
if isinstance(obj, Module):
self.load_module_state_dict(module=obj, state_dict=checkpoint.pop(name), strict=strict)
else:
obj.load_state_dict(checkpoint.pop(name))
else:
state[name] = checkpoint.pop(name)
return checkpoint
[docs] def teardown(self) -> None:
"""This method is called to teardown the training process.
It is the right place to release memory and free other resources.
"""
self.precision.teardown()
assert self.accelerator is not None
self.accelerator.teardown()
self.checkpoint_io.teardown()
[docs] def clip_gradients_norm(
self,
module: torch.nn.Module,
optimizer: Optimizer,
max_norm: Union[float, int],
norm_type: Union[float, int] = 2.0,
error_if_nonfinite: bool = True,
) -> torch.Tensor:
"""Clip gradients by norm."""
self.precision.unscale_gradients(optimizer)
parameters = self.precision.main_params(optimizer)
return torch.nn.utils.clip_grad_norm_(
parameters, max_norm=max_norm, norm_type=norm_type, error_if_nonfinite=error_if_nonfinite
)
[docs] def clip_gradients_value(self, module: torch.nn.Module, optimizer: Optimizer, clip_val: Union[float, int]) -> None:
"""Clip gradients by value."""
self.precision.unscale_gradients(optimizer)
parameters = self.precision.main_params(optimizer)
return torch.nn.utils.clip_grad_value_(parameters, clip_value=clip_val)
@classmethod
def register_strategies(cls, strategy_registry: _StrategyRegistry) -> None:
pass
def _err_msg_joint_setup_required(self) -> str:
return (
f"The `{type(self).__name__}` does not support setting up the module and optimizer(s) independently."
" Please call `setup_module_and_optimizers(model, [optimizer, ...])` to jointly set them up."
)
def _convert_stateful_objects_in_state(
self, state: Dict[str, Union[Module, Optimizer, Any]], filter: Dict[str, Callable[[str, Any], bool]]
) -> Dict[str, Any]:
converted_state: Dict[str, Any] = {}
for key, obj in state.items():
# convert the state
if isinstance(obj, Module):
converted = self.get_module_state_dict(module=obj)
elif isinstance(obj, Optimizer):
converted = self.get_optimizer_state(optimizer=obj)
elif isinstance(obj, _Stateful):
converted = obj.state_dict()
else:
converted = obj
_apply_filter(key, filter, converted, converted_state)
return converted_state
class _BackwardSyncControl(ABC):
"""Interface for any :class:`Strategy` that wants to offer a functionality to enable or disable gradient
synchronization during/after back-propagation.
The most common use-case is gradient accumulation. If a :class:`Strategy` implements this interface, the user can
implement their gradient accumulation loop very efficiently by disabling redundant gradient synchronization.
"""
@abstractmethod
def no_backward_sync(self, module: Module) -> ContextManager:
"""Blocks the synchronization of gradients during the backward pass.
This is a context manager. It is only effective if it wraps a call to `.backward()`.
"""
class _Sharded(ABC):
"""Mixin-interface for any :class:`Strategy` that wants to expose functionality for sharding model parameters."""
@abstractmethod
def module_sharded_context(self) -> ContextManager:
"""A context manager that goes over the instantiation of an :class:`torch.nn.Module` and handles sharding of
parameters on creation.
By sharding layers directly on instantiation, one can reduce peak memory usage and initialization time.
"""
def _validate_keys_for_strict_loading(
requested_keys: Iterable[str], checkpoint_keys: Iterable[str], strict: bool
) -> None:
invalid_keys = [k for k in requested_keys if k not in checkpoint_keys]
if strict and invalid_keys:
raise KeyError(
f"The requested state contains a key '{invalid_keys[0]}' that does not exist in the loaded checkpoint."
f" To disable strict loading, set `strict=False`."
)
def _apply_filter(
key: str, filter: Dict[str, Callable[[str, Any], bool]], source_dict: object, target_dict: Dict[str, Any]
) -> None:
# filter out if necessary
if key in filter and isinstance(source_dict, dict):
filter_fn = filter[key]
for k, v in source_dict.items():
if filter_fn(k, v):
# save the state
target_dict.setdefault(key, {})
target_dict[key][k] = v
else:
# save the state
target_dict[key] = source_dict
