-
Notifications
You must be signed in to change notification settings - Fork 2.3k
/
megatron_gpt_continue_training.py
executable file
·198 lines (174 loc) · 9.5 KB
/
megatron_gpt_continue_training.py
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
# Copyright (c) 2023, 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 os
import tempfile
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_gpt_model import MegatronGPTModel
from nemo.collections.nlp.modules.common.megatron.megatron_init import fake_initialize_model_parallel
from nemo.collections.nlp.parts.nlp_overrides import (
CustomProgressBar,
GradScaler,
MegatronHalfPrecisionPlugin,
NLPDDPStrategy,
NLPSaveRestoreConnector,
PipelineMixedPrecisionPlugin,
)
from nemo.core.config import hydra_runner
from nemo.utils import AppState, logging
from nemo.utils.exp_manager import exp_manager
from nemo.utils.model_utils import inject_model_parallel_rank
def _modify_config(gpt_cfg, cfg, add_cfg_to_tree=False):
"""
This function modifies the original gpt pre-training config (t5_cfg) with attributes from the finetuning config (cfg).
The `add_cfg_to_tree` arg adds `cfg` to the top of the yaml tree which is needed for all `hparams.yaml` files when passed as an arg to `load_from_checkpoint()`.
"""
OmegaConf.set_struct(gpt_cfg, True)
OmegaConf.resolve(cfg)
with open_dict(gpt_cfg):
gpt_cfg.megatron_amp_O2 = cfg.model.get('megatron_amp_O2', False)
gpt_cfg.micro_batch_size = cfg.model.micro_batch_size
gpt_cfg.global_batch_size = cfg.model.global_batch_size
gpt_cfg.sequence_parallel = cfg.model.get("sequence_parallel", False)
gpt_cfg.activations_checkpoint_granularity = cfg.model.get("activations_checkpoint_granularity", None)
gpt_cfg.activations_checkpoint_num_layers = cfg.model.get("activations_checkpoint_num_layers", None)
gpt_cfg.activations_checkpoint_method = cfg.model.get("activations_checkpoint_method", None)
gpt_cfg.data = cfg.model.data
gpt_cfg.optim = cfg.model.optim
gpt_cfg.precision = cfg.trainer.precision
gpt_cfg.restore_from_path = cfg.restore_from_path
gpt_cfg.resume_from_checkpoint = cfg.model.resume_from_checkpoint
gpt_cfg.gradient_as_bucket_view = cfg.model.gradient_as_bucket_view
gpt_cfg.encoder_seq_length = cfg.model.encoder_seq_length
gpt_cfg.max_position_embeddings = cfg.model.max_position_embeddings
gpt_cfg.seq_len_interpolation_factor = cfg.model.seq_len_interpolation_factor
gpt_cfg.use_flash_attention = cfg.model.use_flash_attention
gpt_cfg.tensor_model_parallel_size = cfg.model.get('tensor_model_parallel_size', 1)
gpt_cfg.pipeline_model_parallel_size = cfg.model.get('pipeline_model_parallel_size', 1)
gpt_cfg.pipeline_model_parallel_split_rank = cfg.model.get('pipeline_model_parallel_split_rank', 0)
# This is needed when modifying a hparam file directly to load `.ckpt` files.
# This is not needed to modify the cfg in `.nemo` files.
if add_cfg_to_tree:
OmegaConf.resolve(gpt_cfg)
gpt_cfg.cfg = gpt_cfg
return gpt_cfg
def load_from_nemo(cls, cfg, trainer, gpt_cfg, modify_confg_fn):
gpt_cfg = modify_confg_fn(gpt_cfg, cfg, add_cfg_to_tree=False)
save_restore_connector = NLPSaveRestoreConnector()
if os.path.isdir(cfg.restore_from_path):
save_restore_connector.model_extracted_dir = cfg.restore_from_path
model = cls.restore_from(
restore_path=cfg.restore_from_path,
trainer=trainer,
override_config_path=gpt_cfg,
save_restore_connector=save_restore_connector,
)
return model
def load_from_checkpoint_dir(cls, cfg, trainer, modify_confg_fn):
app_state = AppState()
if cfg.model.tensor_model_parallel_size > 1 or cfg.model.pipeline_model_parallel_size > 1:
app_state.model_parallel_size = cfg.model.tensor_model_parallel_size * cfg.model.pipeline_model_parallel_size
app_state.tensor_model_parallel_size = cfg.model.tensor_model_parallel_size
app_state.pipeline_model_parallel_size = cfg.model.pipeline_model_parallel_size
(
app_state.tensor_model_parallel_rank,
app_state.pipeline_model_parallel_rank,
app_state.model_parallel_size,
app_state.data_parallel_size,
app_state.pipeline_model_parallel_split_rank,
app_state.virtual_pipeline_model_parallel_rank,
) = fake_initialize_model_parallel(
world_size=app_state.model_parallel_size,
rank=trainer.global_rank,
tensor_model_parallel_size_=cfg.model.tensor_model_parallel_size,
pipeline_model_parallel_size_=cfg.model.pipeline_model_parallel_size,
pipeline_model_parallel_split_rank_=cfg.model.pipeline_model_parallel_split_rank,
)
checkpoint_path = inject_model_parallel_rank(
os.path.join(cfg.model.pretrained_checkpoint.checkpoint_dir, cfg.model.pretrained_checkpoint.checkpoint_name)
)
hparams_file = OmegaConf.load(cfg.model.pretrained_checkpoint.hparams_file)
gpt_cfg = modify_confg_fn(hparams_file.cfg, cfg, add_cfg_to_tree=True)
with tempfile.NamedTemporaryFile(suffix='.yaml') as f:
OmegaConf.save(config=gpt_cfg, f=f.name)
model = cls.load_from_checkpoint(checkpoint_path=checkpoint_path, trainer=trainer, hparams_file=f.name,)
return model
def validate_checkpoint_loading_args(cfg):
if cfg.checkpoint_dir is None or not os.path.isdir(cfg.checkpoint_dir):
raise ValueError(f'Checkpoint directory {cfg.checkpoint_dir} does not exist or is not a directory.')
if cfg.checkpoint_name is None:
raise ValueError(f'Checkpoint name {cfg.checkpoint_name} is not valid.')
if cfg.hparams_file is None or not os.path.isfile(cfg.hparams_file):
raise ValueError(f'Hparams file {cfg.hparams_file} does not exist or is not a file.')
@hydra_runner(config_path="conf", config_name="megatron_gpt_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', 'fused_adam') == '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, '16', 'bf16', '16-mixed', 'bf16-mixed']:
scaler = None
if cfg.trainer.precision in [16, '16', '16-mixed']:
scaler = GradScaler(
init_scale=cfg.model.get('native_amp_init_scale', 2 ** 32),
growth_interval=cfg.model.get('native_amp_growth_interval', 1000),
hysteresis=cfg.model.get('hysteresis', 2),
)
plugin_precision = '16-mixed'
else:
plugin_precision = 'bf16-mixed'
if megatron_amp_O2 and not with_distributed_adam:
plugins.append(MegatronHalfPrecisionPlugin(precision=plugin_precision, device='cuda', scaler=scaler))
else:
plugins.append(PipelineMixedPrecisionPlugin(precision=plugin_precision, device='cuda', scaler=scaler))
if cfg.get('cluster_type', None) == 'BCP':
plugins.append(TorchElasticEnvironment())
callbacks = []
# enable_progress_bar is True by default. If cfg.trainer.enable_progress_bar=False, CustomProgressBar is not appended to callbacks
if 'enable_progress_bar' not in cfg.trainer or cfg.trainer.enable_progress_bar:
callbacks.append(CustomProgressBar())
trainer = Trainer(plugins=plugins, strategy=strategy, **cfg.trainer, callbacks=callbacks)
exp_manager(trainer, cfg.exp_manager)
# update resume from checkpoint found by exp_manager
if cfg.model.resume_from_checkpoint is not None:
trainer.ckpt_path = cfg.model.resume_from_checkpoint
logging.info(f'Resuming training from checkpoint: {trainer.ckpt_path}')
if cfg.restore_from_path:
save_restore_connector = NLPSaveRestoreConnector()
if os.path.isdir(cfg.restore_from_path):
save_restore_connector.model_extracted_dir = cfg.restore_from_path
gpt_cfg = MegatronGPTModel.restore_from(
restore_path=cfg.restore_from_path,
trainer=trainer,
return_config=True,
save_restore_connector=save_restore_connector,
)
model = load_from_nemo(MegatronGPTModel, cfg, trainer, gpt_cfg, modify_confg_fn=_modify_config)
elif cfg.model.get("pretrained_checkpoint", None) is not None:
validate_checkpoint_loading_args(cfg.model.pretrained_checkpoint)
model = load_from_checkpoint_dir(MegatronGPTModel, cfg, trainer, modify_confg_fn=_modify_config)
else:
print(' > WARNING: No checkpoint provided. Starting from scratch.')
model = MegatronGPTModel(cfg.model, trainer)
trainer.fit(model)
if __name__ == '__main__':
main()