Source code for pytorch_lightning.strategies.parallel

# 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
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# See the License for the specific language governing permissions and
# limitations under the License.
from abc import ABC, abstractmethod
from contextlib import contextmanager
from typing import Any, Dict, Generator, List, Optional

import torch
from torch import Tensor

import pytorch_lightning as pl
from lightning_lite.plugins import CheckpointIO, ClusterEnvironment
from lightning_lite.utilities.distributed import _all_gather_ddp_if_available, ReduceOp
from pytorch_lightning.plugins import LayerSync
from pytorch_lightning.plugins.precision import PrecisionPlugin
from pytorch_lightning.strategies.strategy import Strategy

[docs]class ParallelStrategy(Strategy, ABC): """Plugin for training with multiple processes in parallel.""" def __init__( self, accelerator: Optional["pl.accelerators.Accelerator"] = None, parallel_devices: Optional[List[torch.device]] = None, cluster_environment: Optional[ClusterEnvironment] = None, checkpoint_io: Optional[CheckpointIO] = None, precision_plugin: Optional[PrecisionPlugin] = None, ): super().__init__(accelerator=accelerator, checkpoint_io=checkpoint_io, precision_plugin=precision_plugin) self.parallel_devices = parallel_devices self.cluster_environment: Optional[ClusterEnvironment] = cluster_environment self._layer_sync: Optional[LayerSync] = None @property @abstractmethod def root_device(self) -> torch.device: """Return the root device.""" @property def global_rank(self) -> int: return self.cluster_environment.global_rank() if self.cluster_environment is not None else 0 @property def local_rank(self) -> int: return self.cluster_environment.local_rank() if self.cluster_environment is not None else 0 @property def node_rank(self) -> int: return self.cluster_environment.node_rank() if self.cluster_environment is not None else 0 @property def world_size(self) -> int: return self.cluster_environment.world_size() if self.cluster_environment is not None else 1 @property def is_global_zero(self) -> bool: return self.global_rank == 0 @property def parallel_devices(self) -> Optional[List[torch.device]]: return self._parallel_devices @parallel_devices.setter def parallel_devices(self, parallel_devices: Optional[List[torch.device]]) -> None: self._parallel_devices = parallel_devices @property def distributed_sampler_kwargs(self) -> Dict[str, Any]: distributed_sampler_kwargs = dict( num_replicas=len(self.parallel_devices) if self.parallel_devices is not None else 0, rank=self.global_rank ) return distributed_sampler_kwargs
[docs] def reconciliate_processes(self, trace: str) -> None: """Function to re-conciliate processes on failure."""
[docs] def all_gather(self, tensor: Tensor, group: Optional[Any] = None, sync_grads: bool = False) -> Tensor: """Perform a all_gather on all processes.""" return _all_gather_ddp_if_available(tensor, group=group, sync_grads=sync_grads)
[docs] def reduce_boolean_decision(self, decision: bool) -> bool: decision = torch.tensor(int(decision), device=self.root_device) decision = self.reduce(decision, reduce_op=ReduceOp.SUM) decision = bool(decision == self.world_size) return decision
[docs] @contextmanager def block_backward_sync(self) -> Generator: """Blocks ddp sync 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, pl.utilities.types.DistributedDataParallel): with self.model.no_sync(): yield None else: yield None
[docs] def teardown(self) -> None: assert self.cluster_environment is not None self.cluster_environment.teardown() super().teardown()

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

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