File size: 7,684 Bytes
b386992 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 | # Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
#
# 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.
import types
from abc import ABC, abstractmethod
from copy import deepcopy
from typing import List, Optional
import lightning.pytorch as L
from lightning.pytorch.utilities.types import OptimizerLRScheduler
from torch.optim import Optimizer
from nemo.lightning.io.mixin import IOMixin
from nemo.lightning.megatron_parallel import CallbackMethods
class LRSchedulerModule(L.Callback, CallbackMethods, IOMixin, ABC):
"""A module to standardize the learning rate scheduler setup and configuration.
This class decouples the learning rate scheduler from the model, similar to how the LightningDataModule
decouples data handling. It also acts as a Callback to hook into the training loop, which can be useful
for adding custom all-reduces, logging, early stopping, etc. Next to that standard Lightning callback-event,
this also supports hooking into the Megatron forward-backward function at a granular level.
Example::
class MyLRSchedulerModule(LRSchedulerModule):
def setup(self, model, optimizer):
# Custom setup logic
...
def scheduler(self, model, optimizers):
# Define and return the learning rate scheduler
...
Methods:
setup(model, optimizer): Sets up the learning rate scheduler.
scheduler(model, optimizers): Abstract method to define the learning rate scheduler.
__call__(model, optimizers): Calls the setup and scheduler methods.
"""
def connect(self, model, optimizer) -> None:
"""Sets up the learning rate scheduler.
Args:
model: The model for which the scheduler is being set up.
optimizer: The optimizer for which the scheduler is being set up.
"""
...
@abstractmethod
def scheduler(self, model, optimizers) -> OptimizerLRScheduler:
"""Abstract method to define the learning rate scheduler.
Args:
model: The model for which the scheduler is being defined.
optimizers: The optimizers for which the scheduler is being defined.
Returns:
OptimizerLRScheduler: The learning rate scheduler.
"""
raise NotImplementedError("The scheduler method should be implemented by subclasses.")
def __call__(self, model, optimizers):
"""Calls the setup and scheduler methods.
Args:
model: The model for which the scheduler is being called.
optimizers: The optimizers for which the scheduler is being called.
Returns:
OptimizerLRScheduler: The learning rate scheduler.
"""
self.connect(model, optimizers)
self._scheduler = self.scheduler(model, optimizers)
if not isinstance(self._scheduler, (dict, tuple)):
return optimizers, self._scheduler
return self._scheduler
class OptimizerModule(L.Callback, CallbackMethods, IOMixin, ABC):
"""A module to standardize the optimizer setup and configuration.
This class decouples the optimizer from the model, similar to how the LightningDataModule
decouples data handling. It also acts as a Callback to hook into the training loop, which can be useful
for adding custom all-reduces, logging, early stopping, etc. Next to that standard Lightning callback-event,
this also supports hooking into the Megatron forward-backward function at a granular level.
Attributes:
lr_scheduler (Optional[LRSchedulerModule]): The learning rate scheduler module.
Example::
class MyOptimizerModule(OptimizerModule):
def __init__(self, lr_scheduler=None):
super().__init__(lr_scheduler)
def setup(self, model):
# Custom setup logic
...
def optimizers(self, model):
# Define and return the optimizers
...
Methods:
connect(model, trainer): Connects the optimizer module to the model and trainer.
setup(model): Sets up the optimizer.
optimizers(model): Abstract method to define the optimizers.
__call__(model, megatron_parallel): Calls the setup and optimizers methods.
"""
def __init__(self, lr_scheduler: Optional[LRSchedulerModule]):
"""Initializes the OptimizerModule.
Args:
lr_scheduler (Optional[LRSchedulerModule]): The learning rate scheduler module.
"""
self.lr_scheduler = lr_scheduler
def connect(self, model: L.LightningModule) -> None:
"""Connects the optimizer module to the model and trainer.
Args:
model (L.LightningModule): The model to which the optimizer module is being connected.
"""
def custom_configure_optimizers(lightning_module_self, megatron_parallel=None):
opt = self(lightning_module_self, megatron_parallel=megatron_parallel)
return opt
model.configure_optimizers = types.MethodType(custom_configure_optimizers, model)
model.optim = self
if hasattr(self, "__io__") and hasattr(model, "__io__"):
if hasattr(model.__io__, "optim"):
model.__io__.optim = deepcopy(self.__io__)
@abstractmethod
def optimizers(self, model) -> List[Optimizer]:
"""Abstract method to define the optimizers.
Args:
model: The model for which the optimizers are being defined.
Returns:
List[Optimizer]: The list of optimizers.
"""
raise NotImplementedError("The optimizers method should be implemented by subclasses.")
def on_train_batch_start(self, trainer, pl_module, batch, batch_idx) -> None:
# pylint: disable=C0116
if self._optimizers is not None:
if len(self._optimizers[0].param_groups) > 0:
lr = self._optimizers[0].param_groups[0]['lr']
else:
lr = 0.0
pl_module.log('lr', lr, batch_size=1, prog_bar=True)
def __call__(self, model: L.LightningModule, megatron_parallel=None) -> OptimizerLRScheduler:
"""Calls the setup and optimizers methods.
Args:
model (L.LightningModule): The model for which the optimizers are being called.
megatron_parallel: Optional parallel model.
Returns:
OptimizerLRScheduler: The optimizers and optionally the learning rate scheduler.
"""
_model = model if megatron_parallel is None else megatron_parallel
callbacks = _model.trainer.callbacks
if self not in callbacks:
callbacks.append(self)
if self.lr_scheduler is not None and self.lr_scheduler not in callbacks:
callbacks.append(self.lr_scheduler)
self._optimizers = self.optimizers(_model)
_opt = self._optimizers[0] if len(self._optimizers) == 1 else self._optimizers
if self.lr_scheduler is not None:
with_scheduler = self.lr_scheduler(_model, _opt)
return with_scheduler
return self._optimizers
|