pytorch_lightning/strategies/horovod.py

# 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 ExitStack
from typing import Any, List, Optional, Tuple, Union

import torch
import torch.nn as nn
from torch.optim import Optimizer
from torch.optim.lr_scheduler import _LRScheduler

import pytorch_lightning as pl
from pytorch_lightning.core.optimizer import LightningOptimizer
from pytorch_lightning.plugins.io.checkpoint_plugin import CheckpointIO
from pytorch_lightning.plugins.precision import PrecisionPlugin
from pytorch_lightning.strategies.parallel import ParallelStrategy
from pytorch_lightning.utilities import _HOROVOD_AVAILABLE
from pytorch_lightning.utilities.distributed import distributed_available
from pytorch_lightning.utilities.distributed import group as dist_group
from pytorch_lightning.utilities.distributed import rank_zero_only, ReduceOp
from pytorch_lightning.utilities.enums import _StrategyType

if _HOROVOD_AVAILABLE:
    import horovod.torch as hvd


class HorovodStrategy(ParallelStrategy):
    """Plugin for Horovod distributed training integration."""

    distributed_backend = _StrategyType.HOROVOD

    def __init__(
        self,
        accelerator: Optional["pl.accelerators.accelerator.Accelerator"] = None,
        parallel_devices: Optional[List[torch.device]] = None,
        checkpoint_io: Optional[CheckpointIO] = None,
        precision_plugin: Optional[PrecisionPlugin] = None,
    ):
        super().__init__(
            accelerator=accelerator,
            parallel_devices=parallel_devices,
            cluster_environment=None,
            checkpoint_io=checkpoint_io,
            precision_plugin=precision_plugin,
        )
        rank_zero_only.rank = self.global_rank
        self._exit_stack: Optional[ExitStack] = None

    @property
    def global_rank(self) -> int:
        return hvd.rank()

    @property
    def local_rank(self) -> int:
        return hvd.local_rank()

    @property
    def world_size(self) -> int:
        return hvd.size()

    @property
    def root_device(self):
        return self.parallel_devices[self.local_rank]

    @property
    def distributed_sampler_kwargs(self):
        distributed_sampler_kwargs = dict(num_replicas=self.world_size, rank=self.global_rank)
        return distributed_sampler_kwargs

    def setup(self, trainer: "pl.Trainer") -> None:
        self.model_to_device()

        super().setup(trainer)

        self._exit_stack = ExitStack()
        self._exit_stack.__enter__()

        if not self.lightning_module.trainer.training:
            # no need to setup optimizers
            return

        def _unpack_lightning_optimizer(opt):
            return opt._optimizer if isinstance(opt, LightningOptimizer) else opt

        optimizers = self.optimizers
        optimizers = [_unpack_lightning_optimizer(opt) for opt in optimizers]

        # Horovod: scale the learning rate by the number of workers to account for
        # increased total batch size
        for optimizer in optimizers:
            for param_group in optimizer.param_groups:
                param_group["lr"] *= self.world_size

        # Horovod: adjust base LR used by schedulers to match scaled optimizer initial LR
        lr_schedulers = self.lightning_module.trainer.lr_schedulers
        for scheduler in lr_schedulers:
            scheduler = scheduler["scheduler"]
            if isinstance(scheduler, _LRScheduler):
                scheduler.base_lrs = [lr * self.world_size for lr in scheduler.base_lrs]

        # Horovod: broadcast parameters & optimizer state to ensure consistent initialization
        hvd.broadcast_parameters(self.lightning_module.state_dict(), root_rank=0)
        for optimizer in optimizers:
            hvd.broadcast_optimizer_state(optimizer, root_rank=0)

        self.optimizers = self._wrap_optimizers(optimizers)
        for optimizer in self.optimizers:
            # Synchronization will be performed explicitly following backward()
            self._exit_stack.enter_context(optimizer.skip_synchronize())

    def barrier(self, *args, **kwargs):
        if distributed_available():
            self.join()

    def broadcast(self, obj: object, src: int = 0) -> object:
        obj = hvd.broadcast_object(obj, src)
        return obj

    def model_to_device(self):
        if self.on_gpu:
            # this can potentially be removed after #8312. Not done due to lack of horovod testing
            torch.cuda.set_device(self.root_device)
        self.model.to(self.root_device)

    def join(self):
        if self.on_gpu:
            hvd.join(self.local_rank)
        else:
            hvd.join()

    def reduce(self, tensor, group: Optional[Any] = None, reduce_op: Optional[Union[ReduceOp, str]] = "mean"):
        """Reduces a tensor from several distributed processes to one aggregated tensor.

        Args:
            tensor: the tensor to sync and reduce
            group: the process group to gather results from. Defaults to all processes (world)
            reduce_op: the reduction operation. Defaults to 'mean'/'avg'.
                Can also be a string 'sum' to calculate the sum during reduction.

        Return:
            reduced value, except when the input was not a tensor the output remains is unchanged
        """
        if group is not None:
            raise ValueError("Horovod does not support allreduce using a subcommunicator at this time. Unset `group`.")

        if reduce_op in (None, "avg", "mean"):
            reduce_op = hvd.Average
        elif reduce_op in ("sum", ReduceOp.SUM):
            reduce_op = hvd.Sum
        else:
            raise ValueError(f"unrecognized `reduce_op`: {reduce_op}")

        # sync all processes before reduction
        self.join()
        return hvd.allreduce(tensor, op=reduce_op)

    def all_gather(
        self, result: torch.Tensor, group: Optional[Any] = dist_group.WORLD, sync_grads: bool = False
    ) -> torch.Tensor:
        if group is not None and group != dist_group.WORLD:
            raise ValueError("Horovod does not support allgather using a subcommunicator at this time. Unset `group`.")

        if len(result.shape) == 0:
            # Convert scalars to single dimension tensors
            result = result.reshape(1)

        # sync and gather all
        self.join()
        return hvd.allgather(result)

    def post_backward(self, closure_loss: torch.Tensor) -> None:
        # synchronize all horovod optimizers.
        for optimizer in self.lightning_module.trainer.optimizers:
            optimizer.synchronize()

    def _wrap_optimizers(self, optimizers: List[Optimizer]) -> List["hvd.DistributedOptimizer"]:
        """Wraps optimizers to perform gradient aggregation via allreduce."""
        return [
            hvd.DistributedOptimizer(opt, named_parameters=self._filter_named_parameters(self.lightning_module, opt))
            if "horovod" not in str(opt.__class__)
            else opt
            for opt in optimizers
        ]

    @staticmethod
    def _filter_named_parameters(model: nn.Module, optimizer: Optimizer) -> List[Tuple[str, nn.Parameter]]:
        opt_params = {p for group in optimizer.param_groups for p in group.get("params", [])}
        return [(name, p) for name, p in model.named_parameters() if p in opt_params]

    def teardown(self) -> None:
        super().teardown()
        self._exit_stack.__exit__(None, None, None)
        self._exit_stack = None
        # Make sure all workers have finished training before returning to the user
        self.join()
        if self.on_gpu:
            # GPU teardown
            self.lightning_module.cpu()
            # clean up memory
            torch.cuda.empty_cache()