Source code for pytorch_lightning.plugins.precision.deepspeed_precision
# 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 typing import Any, Callable, Union
from torch import Tensor
from torch.nn import Module
from torch.optim import LBFGS, Optimizer
import pytorch_lightning as pl
from pytorch_lightning.plugins.precision.precision_plugin import PrecisionPlugin
from pytorch_lightning.utilities import GradClipAlgorithmType
from pytorch_lightning.utilities.exceptions import MisconfigurationException
from pytorch_lightning.utilities.imports import _DEEPSPEED_AVAILABLE
from pytorch_lightning.utilities.model_helpers import is_overridden
from pytorch_lightning.utilities.warnings import WarningCache
if _DEEPSPEED_AVAILABLE:
from deepspeed import DeepSpeedEngine
warning_cache = WarningCache()
[docs]class DeepSpeedPrecisionPlugin(PrecisionPlugin):
"""Precision plugin for DeepSpeed integration."""
def __init__(self, precision: int) -> None:
super().__init__()
self.precision = precision
[docs] def backward(self, model: "pl.LightningModule", closure_loss: Tensor, *args: Any, **kwargs: Any) -> None:
if is_overridden("backward", model):
warning_cache.warn(
"You have overridden the `LightningModule.backward` hook but it will be ignored since DeepSpeed handles"
" the backward logic internally."
)
deepspeed_engine: DeepSpeedEngine = model.trainer.model
deepspeed_engine.backward(closure_loss, *args, **kwargs)
def _run_backward(self, tensor: Tensor, model: Module, *args: Any, **kwargs: Any) -> None:
model.backward(tensor, *args, **kwargs)
[docs] def optimizer_step(
self,
model: Union["pl.LightningModule", Module],
optimizer: Optimizer,
optimizer_idx: int,
closure: Callable[[], Any],
**kwargs: Any,
) -> None:
if isinstance(optimizer, LBFGS):
raise MisconfigurationException(
f"DeepSpeed and the LBFGS optimizer are not compatible (optimizer {optimizer_idx})."
)
closure_result = closure()
self._after_closure(model, optimizer, optimizer_idx)
skipped_backward = closure_result is None
# in manual optimization, the closure does not return a value
if isinstance(model, pl.LightningModule) and model.automatic_optimization and skipped_backward:
raise MisconfigurationException(
"Skipping backward by returning `None` from your `training_step` is not supported by `DeepSpeed`"
)
# DeepSpeed handles the optimizer step internally
deepspeed_engine = model.trainer.model if isinstance(model, pl.LightningModule) else model
deepspeed_engine.step(**kwargs)
[docs] def clip_gradients(
self,
optimizer: Optimizer,
clip_val: Union[int, float] = 0.0,
gradient_clip_algorithm: GradClipAlgorithmType = GradClipAlgorithmType.NORM,
) -> None:
"""DeepSpeed handles gradient clipping internally."""
def _track_grad_norm(self, trainer: "pl.Trainer") -> None:
if trainer.track_grad_norm == -1:
return
# the gradients are not available in the model due to gradient partitioning in zero stage >= 2
warning_cache.warn(
f"You set `Trainer(track_grad_norm={trainer.track_grad_norm!r})' but this is not supported for DeepSpeed."
" The setting will be ignored."
)