-
Notifications
You must be signed in to change notification settings - Fork 3.4k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Use fsdp module to initialize precision scalar for fsdp native (#14092)
Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com> Co-authored-by: Laverne Henderson <laverne.henderson@coupa.com> Co-authored-by: Rohit Gupta <rohitgr1998@gmail.com>
- Loading branch information
1 parent
d03a7e9
commit 9f189ff
Showing
10 changed files
with
110 additions
and
44 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
65 changes: 65 additions & 0 deletions
65
src/pytorch_lightning/plugins/precision/fsdp_native_native_amp.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,65 @@ | ||
# 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, Optional, Union | ||
|
||
import torch | ||
|
||
from pytorch_lightning.plugins.precision.native_amp import NativeMixedPrecisionPlugin | ||
from pytorch_lightning.utilities.enums import PrecisionType | ||
from pytorch_lightning.utilities.exceptions import MisconfigurationException | ||
from pytorch_lightning.utilities.imports import _TORCH_GREATER_EQUAL_1_12 | ||
|
||
if _TORCH_GREATER_EQUAL_1_12: | ||
from torch.distributed.fsdp.fully_sharded_data_parallel import MixedPrecision | ||
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler | ||
else: | ||
MixedPrecision = None # type: ignore[misc,assignment] | ||
|
||
|
||
class FullyShardedNativeNativeMixedPrecisionPlugin(NativeMixedPrecisionPlugin): | ||
"""Native AMP for Fully Sharded Native Training.""" | ||
|
||
def __init__( | ||
self, precision: Union[str, int], device: str, scaler: Optional[torch.cuda.amp.GradScaler] = None | ||
) -> None: | ||
if not _TORCH_GREATER_EQUAL_1_12: | ||
raise MisconfigurationException( | ||
"`FullyShardedNativeNativeMixedPrecisionPlugin` is supported from PyTorch v1.12.0 onwards." | ||
) | ||
super().__init__(precision, device, scaler=ShardedGradScaler() if scaler is None and precision == 16 else None) | ||
|
||
def clip_grad_by_norm(self, *_: Any, **__: Any) -> None: | ||
# see https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.FullyShardedDataParallel.clip_grad_norm_ | ||
# section `Gradient Clipping`, using `torch.nn.utils.clip_grad_norm_` is incorrect | ||
# for FSDP module. To overcome this, needs to call sharded_module.clip_grad_norm(clip_val) | ||
# however we rely on LightningModule's configure_sharded_model to wrap FSDP, it would be hard to | ||
# trace back the root FSDP. Now we only support clip by value. | ||
raise MisconfigurationException( | ||
f"`gradient_clip_algorithm='norm'` is currently not supported for `{self.__class__.__name__}`" | ||
) | ||
|
||
@property | ||
def mixed_precision_config(self) -> Optional[MixedPrecision]: | ||
assert MixedPrecision is not None | ||
if self.precision == PrecisionType.HALF: | ||
dtype = torch.float16 | ||
elif self.precision == PrecisionType.BFLOAT: | ||
dtype = torch.bfloat16 | ||
else: | ||
raise MisconfigurationException(f"Was unable to infer precision type, received {self.precision!r}.") | ||
return MixedPrecision( | ||
param_dtype=dtype, | ||
reduce_dtype=dtype, | ||
buffer_dtype=dtype, | ||
) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters