NeMo / examples /nlp /language_modeling /megatron_bert_pretraining.py
camenduru's picture
thanks to NVIDIA ❤
7934b29
# Copyright (c) 2021, 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.
from omegaconf.omegaconf import OmegaConf, open_dict
from pytorch_lightning import Trainer
from pytorch_lightning.plugins.environments import TorchElasticEnvironment
from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
from nemo.collections.nlp.models.language_modeling.megatron_bert_model import MegatronBertModel
from nemo.collections.nlp.parts.nlp_overrides import (
GradScaler,
MegatronHalfPrecisionPlugin,
NLPDDPStrategy,
PipelineMixedPrecisionPlugin,
)
from nemo.core.config import hydra_runner
from nemo.utils import logging
from nemo.utils.exp_manager import exp_manager
@hydra_runner(config_path="conf", config_name="megatron_bert_config")
def main(cfg) -> None:
logging.info("\n\n************** Experiment configuration ***********")
logging.info(f'\n{OmegaConf.to_yaml(cfg)}')
megatron_amp_o2 = cfg.model.get('megatron_amp_O2', False)
with_distributed_adam = cfg.model.optim.get('name') == 'distributed_fused_adam'
plugins = []
strategy = NLPDDPStrategy(
no_ddp_communication_hook=True,
gradient_as_bucket_view=cfg.model.gradient_as_bucket_view,
find_unused_parameters=False,
)
if cfg.trainer.precision in [16, 'bf16']:
scaler = None
if cfg.trainer.precision == 16:
scaler = GradScaler(
init_scale=cfg.model.get('native_amp_init_scale', 2 ** 32),
growth_interval=cfg.model.get('native_amp_growth_interval', 1000),
)
if megatron_amp_o2 and not with_distributed_adam:
plugins.append(MegatronHalfPrecisionPlugin(precision=cfg.trainer.precision, device='cuda', scaler=scaler))
else:
plugins.append(PipelineMixedPrecisionPlugin(precision=cfg.trainer.precision, device='cuda', scaler=scaler))
if cfg.get('cluster_type', None) == 'BCP':
plugins.append(TorchElasticEnvironment())
trainer = Trainer(plugins=plugins, strategy=strategy, **cfg.trainer)
exp_manager(trainer, cfg.exp_manager)
# update resume from checkpoint found by exp_manager
resume_from_checkpoint = trainer._checkpoint_connector.resume_from_checkpoint_fit_path
# resume_from_checkpoint = uninject_model_parallel_rank(resume_from_checkpoint)
logging.info(f'Resuming training from checkpoint: {resume_from_checkpoint}')
trainer._checkpoint_connector = CheckpointConnector(trainer, resume_from_checkpoint=resume_from_checkpoint)
# hydra interpolation does not work here as the interpolation key is lost when PTL saves hparams
with open_dict(cfg):
cfg.model.precision = cfg.trainer.precision
model = MegatronBertModel(cfg.model, trainer)
trainer.fit(model)
if __name__ == '__main__':
main()