Source code for lightning.pytorch.plugins.precision.deepspeed
# Copyright The Lightning AI 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, cast, Literal, Optional, TYPE_CHECKING, Union
from torch import Tensor
from torch.optim import LBFGS, Optimizer
from typing_extensions import get_args
import lightning.pytorch as pl
from lightning.fabric.strategies.deepspeed import _DEEPSPEED_AVAILABLE
from lightning.fabric.utilities.types import Steppable
from lightning.pytorch.plugins.precision.precision_plugin import PrecisionPlugin
from lightning.pytorch.utilities import GradClipAlgorithmType
from lightning.pytorch.utilities.exceptions import MisconfigurationException
from lightning.pytorch.utilities.model_helpers import is_overridden
from lightning.pytorch.utilities.rank_zero import WarningCache
if TYPE_CHECKING and _DEEPSPEED_AVAILABLE:
import deepspeed
warning_cache = WarningCache()
_PRECISION_INPUT = Literal["32-true", "16-mixed", "bf16-mixed"]
[docs]class DeepSpeedPrecisionPlugin(PrecisionPlugin):
"""Precision plugin for DeepSpeed integration.
.. warning:: This is an :ref:`experimental <versioning:Experimental API>` feature.
Args:
precision: Full precision (32), half precision (16) or bfloat16 precision (bf16).
Raises:
ValueError:
If unsupported ``precision`` is provided.
"""
def __init__(self, precision: Literal["32-true", "16-mixed", "bf16-mixed"]) -> None:
supported_precision = get_args(_PRECISION_INPUT)
if precision not in supported_precision:
raise ValueError(
f"`Trainer(strategy='deepspeed', precision={precision!r})` is not supported."
f" `precision` must be one of: {supported_precision}."
)
self.precision = cast(_PRECISION_INPUT, str(precision))
[docs] def backward( # type: ignore[override]
self,
tensor: Tensor,
model: "pl.LightningModule",
optimizer: Optional[Steppable],
*args: Any,
**kwargs: Any,
) -> None:
r"""Performs back-propagation using DeepSpeed's engine.
Args:
tensor: the loss tensor
model: the model to be optimized
optimizer: ignored for DeepSpeed
\*args: additional positional arguments for the :meth:`deepspeed.DeepSpeedEngine.backward` call
\**kwargs: additional keyword arguments for the :meth:`deepspeed.DeepSpeedEngine.backward` call
"""
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: "deepspeed.DeepSpeedEngine" = model.trainer.model
deepspeed_engine.backward(tensor, *args, **kwargs)
[docs] def optimizer_step( # type: ignore[override]
self,
optimizer: Steppable,
model: "pl.LightningModule",
closure: Callable[[], Any],
**kwargs: Any,
) -> Any:
if isinstance(optimizer, LBFGS):
raise MisconfigurationException("DeepSpeed and the LBFGS optimizer are not compatible.")
closure_result = closure()
self._after_closure(model, optimizer)
skipped_backward = closure_result is None
# in manual optimization, the closure does not return a value
if 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: "deepspeed.DeepSpeedEngine" = model.trainer.model
return 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."""