Shortcuts

Source code for pytorch_lightning.strategies.sharded

# Copyright The PyTorch Lightning 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.
from contextlib import contextmanager
from typing import Any, Dict, Generator, List, Tuple

from torch import Tensor
from torch.nn import Module
from torch.optim import Optimizer

import pytorch_lightning as pl
from lightning_fabric.utilities.optimizer import _optimizers_to_device
from pytorch_lightning.core.optimizer import LightningOptimizer
from pytorch_lightning.overrides.base import _LightningModuleWrapperBase, _LightningPrecisionModuleWrapperBase
from pytorch_lightning.overrides.fairscale import _FAIRSCALE_AVAILABLE, _reinit_optimizers_with_oss
from pytorch_lightning.strategies.ddp import DDPStrategy
from pytorch_lightning.trainer.states import TrainerFn
from pytorch_lightning.utilities.exceptions import MisconfigurationException
from pytorch_lightning.utilities.rank_zero import rank_zero_deprecation

if _FAIRSCALE_AVAILABLE:
    from fairscale.nn.data_parallel.sharded_ddp import ShardedDataParallel
    from fairscale.optim import OSS
else:
    OSS = ShardedDataParallel = object


[docs]class DDPShardedStrategy(DDPStrategy): """Optimizer and gradient sharded training provided by FairScale.""" strategy_name = "ddp_sharded" _REDUCE_BUFFER_SIZE_DEFAULT: int = 2**23 # 8M def __init__(self, *args: Any, **kwargs: Any) -> None: rank_zero_deprecation( "PyTorch Lightning's sharded implementation using FairScale has been deprecated in v1.9.0 and will be" " removed in v2.0.0. You can try using the `Trainer(strategy='fsdp_native')` instead." " The difference is that native FSDP uses PyTorch's implementation and the current strategy uses" " FairScale's implementation (which was upstreamed to PyTorch). After removal, `strategy='fsdp'` will use" " the native version by default." ) super().__init__(*args, **kwargs)
[docs] def connect(self, model: "pl.LightningModule") -> None: if not _FAIRSCALE_AVAILABLE: # pragma: no cover raise MisconfigurationException( "`DDPShardedStrategy` requires `fairscale` to be installed." " Install it by running `pip install fairscale`." ) return super().connect(model)
[docs] def setup(self, trainer: "pl.Trainer") -> None: # share ddp pids to all processes self._rank_0_will_call_children_scripts: bool = self.broadcast(self._rank_0_will_call_children_scripts) if self._should_run_deadlock_detection(): self._share_information_to_prevent_deadlock() assert self.accelerator is not None self.accelerator.setup(trainer) # move the model to the correct device self.model_to_device() # skip wrapping the model if we are not fitting as no gradients need to be exchanged trainer_fn = trainer.state.fn if trainer_fn == TrainerFn.FITTING: if self._layer_sync: assert self.model is not None self.model = self._layer_sync.apply(self.model) self.setup_precision_plugin() if trainer_fn == TrainerFn.FITTING: self.configure_ddp()
def configure_ddp(self) -> None: self._set_ddp_kwargs() assert self.lightning_module is not None self.setup_optimizers(self.lightning_module.trainer) assert isinstance(self.model, (pl.LightningModule, _LightningPrecisionModuleWrapperBase)) self.model, self.optimizers = self._setup_model_and_optimizers( model=_LightningModuleWrapperBase(self.model), optimizers=self.optimizers, ) _optimizers_to_device(self.optimizers, self.root_device) def _set_ddp_kwargs(self) -> None: if "reduce_buffer_size" not in self._ddp_kwargs: # For multi-node training, enabling bucketing will improve performance. self._ddp_kwargs["reduce_buffer_size"] = self._REDUCE_BUFFER_SIZE_DEFAULT if self.num_nodes > 1 else 0 def _setup_model_and_optimizers(self, model: Module, optimizers: List[Optimizer]) -> Tuple[Module, List[Optimizer]]: """Wraps the model and optimizers with fairscale components. Return: The model wrapped into a :class:`~fairscale.nn.data_parallel.ShardedDataParallel` module and a list of optimizer wrapped in :class:~`fairscale.optim.OSS`. """ optimizers = self._wrap_optimizers(optimizers) model = ShardedDataParallel(model, sharded_optimizer=optimizers, **self._ddp_kwargs) return model, optimizers def _wrap_optimizers(self, optimizers: List[Optimizer]) -> List["OSS"]: assert self.lightning_module is not None if self.model is not None and self.lightning_module.trainer.state.fn != TrainerFn.FITTING: return optimizers optimizers = [o._optimizer if isinstance(o, LightningOptimizer) else o for o in optimizers] return _reinit_optimizers_with_oss(optimizers, self.precision_plugin, self.num_nodes)
[docs] def pre_backward(self, closure_loss: Tensor) -> None: pass
[docs] @contextmanager def block_backward_sync(self) -> Generator: """Blocks syncing gradients behaviour on backwards pass. This is useful for skipping sync when accumulating gradients, reducing communication overhead Returns: context manager with sync behaviour off """ if isinstance(self.model, ShardedDataParallel): with self.model.no_sync(): yield None else: yield None
def post_training_step(self) -> None: pass @classmethod def register_strategies(cls, strategy_registry: Dict) -> None: strategy_registry.register( "ddp_sharded_find_unused_parameters_false", cls, description="DDP Sharded Strategy with `find_unused_parameters` as False", find_unused_parameters=False, ) strategy_registry.register( cls.strategy_name, cls, description=f"{cls.__class__.__name__}", )

© Copyright Copyright (c) 2018-2023, Lightning AI et al...

Built with Sphinx using a theme provided by Read the Docs.