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 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_lite.strategies.fairscale import _FAIRSCALE_AVAILABLE, _reinit_optimizers_with_oss
from lightning_lite.utilities.optimizer import _optimizers_to_device
from pytorch_lightning.core.optimizer import LightningOptimizer
from pytorch_lightning.overrides.base import _LightningModuleWrapperBase, _LightningPrecisionModuleWrapperBase
from pytorch_lightning.strategies.ddp import DDPStrategy
from pytorch_lightning.trainer.states import TrainerFn
from pytorch_lightning.utilities.exceptions import MisconfigurationException

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


class DDPShardedStrategy(DDPStrategy):
    """Optimizer and gradient sharded training provided by FairScale."""

    strategy_name = "ddp_sharded"
    _REDUCE_BUFFER_SIZE_DEFAULT: int = 2**23  # 8M

    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)

    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)

    def pre_backward(self, closure_loss: Tensor) -> None:
        pass

    @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__}",
        )