arguments.py 16.3 KB
Newer Older
Raul Puri's avatar
Raul Puri committed
1
# coding=utf-8
Mohammad's avatar
Mohammad committed
2
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
Raul Puri's avatar
Raul Puri committed
3
4
5
6
7
8
9
10
11
12
13
14
15
#
# 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.

Mohammad's avatar
Mohammad committed
16
"""Megatron arguments."""
Raul Puri's avatar
Raul Puri committed
17
18
19
20
21

import argparse
import os


Mohammad's avatar
Mohammad committed
22
23
24
def parse_args(extra_args_provider=None, defaults={}):
    """Parse all arguments."""
    parser = argparse.ArgumentParser(description='Megatron-LM Arguments')
Mohammad's avatar
Mohammad committed
25

Mohammad's avatar
Mohammad committed
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
    # Standard arguments.
    parser = _add_network_size_args(parser)
    parser = _add_regularization_args(parser)
    parser = _add_training_args(parser)
    parser = _add_initialization_args(parser)
    parser = _add_learning_rate_args(parser)
    parser = _add_checkpointing_args(parser)
    parser = _add_mixed_precision_args(parser)
    parser = _add_distributed_args(parser)
    parser = _add_validation_args(parser)
    parser = _add_data_args(parser)
    parser = _add_autoresume_args(parser)

    # Custom arguments.
    if extra_args_provider is not None:
        parser = extra_args_provider(parser)
Mohammad's avatar
Mohammad committed
42

Mohammad's avatar
Mohammad committed
43
44
    # Parse.
    args = parser.parse_args()
Mohammad's avatar
Mohammad committed
45

Mohammad's avatar
Mohammad committed
46
47
    # Set input defaults.
    for key in defaults:
Mohammad's avatar
Mohammad committed
48
49
50
51
52
        # For default to be valid, it should not be provided in the
        # arguments that are passed to the program. We check this by
        # ensuring the arg is set to None.
        assert getattr(args, key) is None, \
            'defaults can only be overwritten for args with None values.'
Mohammad's avatar
Mohammad committed
53
        setattr(args, key, defaults[key])
Mohammad's avatar
Mohammad committed
54

55
    # Check required arguments.
Mohammad's avatar
Mohammad committed
56
57
58
59
    required_args = ['num_layers', 'hidden_size', 'num_attention_heads',
                     'max_position_embeddings']
    for req_arg in required_args: 
        _check_arg_is_not_none(args, req_arg)
60

Mohammad's avatar
Mohammad committed
61
62
63
64
65
66
67
    # Distributed args.
    args.rank = int(os.getenv('RANK', '0'))
    args.world_size = int(os.getenv("WORLD_SIZE", '1'))
    args.model_parallel_size = min(args.model_parallel_size, args.world_size)
    if args.rank == 0:
        print('using world size: {} and model-parallel size: {} '.format(
            args.world_size, args.model_parallel_size))
Mohammad's avatar
Mohammad committed
68

Mohammad's avatar
Mohammad committed
69
70
71
72
    # Fp16 loss scaling.
    args.dynamic_loss_scale = False
    if args.loss_scale is None:
        args.dynamic_loss_scale = True
Mohammad's avatar
Mohammad committed
73

Mohammad's avatar
Mohammad committed
74
75
    # Checks.
    assert args.hidden_size % args.num_attention_heads == 0
Mohammad's avatar
Mohammad committed
76
77
78
79
    if args.seq_length is not None:
        assert args.max_position_embeddings >= args.seq_length
    if args.lr is not None:
        assert args.min_lr <= args.lr
Mohammad's avatar
Mohammad committed
80
81
    if args.save is not None:
        assert args.save_interval is not None
Mohammad's avatar
Mohammad committed
82

Mohammad's avatar
Mohammad committed
83
84
    _print_args(args)
    return args
Mohammad's avatar
Mohammad committed
85
86


Mohammad's avatar
Mohammad committed
87
88
89
90
91
92
93
94
95
96
97
def _print_args(args):
    """Print arguments."""
    if args.rank == 0:
        print('-------------------- arguments --------------------', flush=True)
        str_list = []
        for arg in vars(args):
            dots = '.' * (32 - len(arg))
            str_list.append('  {} {} {}'.format(arg, dots, getattr(args, arg)))
        for arg in sorted(str_list, key=lambda x: x.lower()):
            print(arg, flush=True)
        print('---------------- end of arguments ----------------', flush=True)
Mohammad's avatar
Mohammad committed
98
99


100
101
102
103
def _check_arg_is_not_none(args, arg):
    assert getattr(args, arg) is not None, '{} argument is None'.format(arg)


Mohammad's avatar
Mohammad committed
104
def _add_network_size_args(parser):
Mohammad's avatar
Mohammad committed
105
    group = parser.add_argument_group(title='network size')
Mohammad's avatar
Mohammad committed
106

107
    group.add_argument('--num-layers', type=int, default=None,
Mohammad's avatar
Mohammad committed
108
                       help='Number of transformer layers.')
109
    group.add_argument('--hidden-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
110
                       help='Tansformer hidden size.')
111
    group.add_argument('--num-attention-heads', type=int, default=None,
Mohammad's avatar
Mohammad committed
112
                       help='Number of transformer attention heads.')
113
    group.add_argument('--max-position-embeddings', type=int, default=None,
Mohammad's avatar
Mohammad committed
114
115
116
117
118
                       help='Maximum number of position embeddings to use. '
                       'This is the size of position embedding.')
    group.add_argument('--make-vocab-size-divisible-by', type=int, default=128,
                       help='Pad the vocab size to be divisible by this value.'
                       'This is added for computational efficieny reasons.')
Mohammad's avatar
Mohammad committed
119
120
    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
                       help='Layer norm epsilon.')
Mohammad's avatar
Mohammad committed
121
122
123
124
    group.add_argument('--apply-residual-connection-post-layernorm',
                       action='store_true',
                       help='If set, use original BERT residula connection '
                       'ordering.')
125
126
127
128
    group.add_argument('--openai-gelu', action='store_true',
                       help='Use OpenAIs GeLU implementation. This option'
                       'should not be used unless for backward compatibility'
                       'reasons.')
Mohammad's avatar
Mohammad committed
129

Mohammad's avatar
Mohammad committed
130
131
132
    return parser


Mohammad's avatar
Mohammad committed
133
def _add_regularization_args(parser):
Mohammad's avatar
Mohammad committed
134
135
136
137
138
139
140
141
142
143
144
145
146
    group = parser.add_argument_group(title='regularization')

    group.add_argument('--attention-dropout', type=float, default=0.1,
                       help='Post attention dropout ptobability.')
    group.add_argument('--hidden-dropout', type=float, default=0.1,
                       help='Dropout probability for hidden state transformer.')
    group.add_argument('--weight-decay', type=float, default=0.01,
                       help='Weight decay coefficient for L2 regularization.')
    group.add_argument('--clip-grad', type=float, default=1.0,
                       help='Gradient clipping based on global L2 norm.')

    return parser

Mohammad's avatar
Mohammad committed
147
148

def _add_training_args(parser):
Mohammad's avatar
Mohammad committed
149
150
    group = parser.add_argument_group(title='training')

Mohammad's avatar
Mohammad committed
151
    group.add_argument('--batch-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
152
153
154
155
156
157
158
159
                       help='Batch size per model instance (local batch size). '
                       'Global batch size is local batch size times data '
                       'parallel size.')
    group.add_argument('--checkpoint-activations', action='store_true',
                       help='Checkpoint activation to allow for training '
                       'with larger models, sequences, and batch sizes.')
    group.add_argument('--checkpoint-num-layers', type=int, default=1,
                       help='chunk size (number of layers) for checkpointing.')
Mohammad's avatar
Mohammad committed
160
    group.add_argument('--train-iters', type=int, default=None,
Mohammad's avatar
Mohammad committed
161
162
163
164
165
166
167
168
169
170
171
172
173
                       help='Total number of iterations to train over all '
                       'training runs.')
    group.add_argument('--log-interval', type=int, default=100,
                       help='Report loss and timing interval.')
    group.add_argument('--exit-interval', type=int, default=None,
                       help='Exit the program after the iteration is divisible '
                       'by this value.')
    group.add_argument('--tensorboard-dir', type=str, default=None,
                       help='Write TensorBoard logs to this directory.')

    return parser


Mohammad's avatar
Mohammad committed
174
def _add_initialization_args(parser):
Mohammad's avatar
Mohammad committed
175
176
177
178
179
180
181
182
    group = parser.add_argument_group(title='initialization')

    group.add_argument('--seed', type=int, default=1234,
                       help='Random seed used for python, numpy, '
                       'pytorch, and cuda.')
    group.add_argument('--init-method-std', type=float, default=0.02,
                       help='Standard deviation of the zero mean normal '
                       'distribution used for weight initialization.')
Mohammad's avatar
Mohammad committed
183

Mohammad's avatar
Mohammad committed
184
185
186
    return parser


Mohammad's avatar
Mohammad committed
187
def _add_learning_rate_args(parser):
Mohammad's avatar
Mohammad committed
188
189
    group = parser.add_argument_group(title='learning rate')

Mohammad's avatar
Mohammad committed
190
    group.add_argument('--lr', type=float, default=None,
Mohammad's avatar
Mohammad committed
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
                       help='Initial learning rate. Depending on decay style '
                       'and initial warmup, the learing rate at each '
                       'iteration would be different.')
    group.add_argument('--lr-decay-style', type=str, default='linear',
                       choices=['constant', 'linear', 'cosine', 'exponential'],
                       help='Learning rate decay function.')
    group.add_argument('--lr-decay-iters', type=int, default=None,
                       help='number of iterations to decay learning rate over,'
                       ' If None defaults to `--train-iters`')
    group.add_argument('--min-lr', type=float, default=0.0,
                       help='Minumum value for learning rate. The scheduler'
                       'clip values below this threshold.')
    group.add_argument('--warmup', type=float, default=0.01,
                       help='Percentage of total iterations to warmup on '
                       '(.01 = 1 percent of all training iters).')
    group.add_argument('--override-lr-scheduler', action='store_true',
                       help='Reset the values of the scheduler (learning rate,'
                       'warmup iterations, minimum learning rate, maximum '
                       'number of iterations, and decay style from input '
                       'arguments and ignore values from checkpoints. Note'
                       'that all the above values will be reset.')
    group.add_argument('--use-checkpoint-lr-scheduler', action='store_true',
                       help='Use checkpoint to set the values of the scheduler '
                       '(learning rate, warmup iterations, minimum learning '
                       'rate, maximum number of iterations, and decay style '
                       'from checkpoint and ignore input arguments.')

    return parser


Mohammad's avatar
Mohammad committed
221
def _add_checkpointing_args(parser):
Mohammad's avatar
Mohammad committed
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
    group = parser.add_argument_group(title='checkpointing')

    group.add_argument('--save', type=str, default=None,
                       help='Output directory to save checkpoints to.')
    group.add_argument('--save-interval', type=int, default=None,
                       help='Number of iterations between checkpoint saves.')
    group.add_argument('--no-save-optim', action='store_true',
                       help='Do not save current optimizer.')
    group.add_argument('--no-save-rng', action='store_true',
                       help='Do not save current rng state.')
    group.add_argument('--load', type=str, default=None,
                       help='Directory containing a model checkpoint.')
    group.add_argument('--no-load-optim', action='store_true',
                       help='Do not load optimizer when loading checkpoint.')
    group.add_argument('--no-load-rng', action='store_true',
                       help='Do not load rng state when loading checkpoint.')
    group.add_argument('--finetune', action='store_true',
                       help='Load model for finetuning. Do not load optimizer '
                       'or rng state from checkpoint and set iteration to 0. '
                       'Assumed when loading a release checkpoint.')

    return parser


Mohammad's avatar
Mohammad committed
246
def _add_mixed_precision_args(parser):
Mohammad's avatar
Mohammad committed
247
248
249
250
251
252
253
254
255
256
    group = parser.add_argument_group(title='mixed precision')

    group.add_argument('--fp16', action='store_true',
                       help='Run model in fp16 mode.')
    group.add_argument('--apply-query-key-layer-scaling', action='store_true',
                       help='Scale Q * K^T by 1 / layer-number. If this flag '
                       'is set, then it will automatically set '
                       'attention-softmax-in-fp32 to true')
    group.add_argument('--attention-softmax-in-fp32', action='store_true',
                       help='Run attention masking and softmax in fp32.')
Mohammad's avatar
Mohammad committed
257
258
    group.add_argument('--fp32-allreduce', action='store_true',
                       help='All-reduce in fp32')
Mohammad's avatar
Mohammad committed
259
260
261
262
263
264
265
266
267
268
269
270
271
272
    group.add_argument('--hysteresis', type=int, default=2,
                       help='hysteresis for dynamic loss scaling')
    group.add_argument('--loss-scale', type=float, default=None,
                       help='Static loss scaling, positive power of 2 '
                       'values can improve fp16 convergence. If None, dynamic'
                       'loss scaling is used.')
    group.add_argument('--loss-scale-window', type=float, default=1000,
                       help='Window over which to raise/lower dynamic scale.')
    group.add_argument('--min-scale', type=float, default=1,
                       help='Minimum loss scale for dynamic loss scale.')

    return parser


Mohammad's avatar
Mohammad committed
273
def _add_distributed_args(parser):
Mohammad's avatar
Mohammad committed
274
275
    group = parser.add_argument_group(title='mixed precision')

Mohammad's avatar
Mohammad committed
276
277
    group.add_argument('--model-parallel-size', type=int, default=1,
                       help='Size of the model parallel.')
Mohammad's avatar
Mohammad committed
278
279
280
281
    group.add_argument('--distributed-backend', default='nccl',
                       choices=['nccl', 'gloo'],
                       help='Which backend to use for distributed training.')
    group.add_argument('--DDP-impl', default='local',
Mohammad's avatar
Mohammad committed
282
                       choices=['local', 'torch'],
Mohammad's avatar
Mohammad committed
283
284
285
286
287
288
289
290
                       help='which DistributedDataParallel implementation '
                       'to use.')
    group.add_argument('--local_rank', type=int, default=None,
                       help='local rank passed from distributed launcher.')

    return parser


Mohammad's avatar
Mohammad committed
291
def _add_validation_args(parser):
Mohammad's avatar
Mohammad committed
292
293
294
295
296
297
298
299
300
    group = parser.add_argument_group(title='validation')

    group.add_argument('--eval-iters', type=int, default=100,
                       help='Number of iterations to run for evaluation'
                       'validation/test for.')
    group.add_argument('--eval-interval', type=int, default=1000,
                       help='Interval between running evaluation on '
                       'validation set.')

Mohammad's avatar
Mohammad committed
301
302
303
    return parser


Mohammad's avatar
Mohammad committed
304
def _add_data_args(parser):
Mohammad's avatar
Mohammad committed
305
306
    group = parser.add_argument_group(title='data and dataloader')

Mohammad's avatar
Mohammad committed
307
    group.add_argument('--data-path', type=str, default=None,
Mohammad's avatar
Mohammad committed
308
                       help='Path to combined dataset to split.')
Mohammad's avatar
Mohammad committed
309
    group.add_argument('--split', type=str, default='969, 30, 1',
Mohammad's avatar
Mohammad committed
310
311
312
313
                       help='Comma-separated list of proportions for training,'
                       ' validation, and test split. For example the split '
                       '`90,5,5` will use 90% of data for training, 5% for '
                       'validation and 5% for test.')
Mohammad's avatar
Mohammad committed
314
    group.add_argument('--vocab-file', type=str, default=None,
Mohammad's avatar
Mohammad committed
315
                       help='Path to the vocab file.')
Mohammad's avatar
Mohammad committed
316
317
    group.add_argument('--merge-file', type=str, default=None,
                       help='Path to the BPE merge file.')
Mohammad's avatar
Mohammad committed
318
    group.add_argument('--seq-length', type=int, default=None,
Mohammad's avatar
Mohammad committed
319
320
321
322
323
324
325
326
327
                       help="Maximum sequence length to process.")
    group.add_argument('--mask-prob', type=float, default=0.15,
                       help='Probability of replacing a token with mask.')
    group.add_argument('--short-seq-prob', type=float, default=0.1,
                       help='Probability of producing a short sequence.')
    group.add_argument('--mmap-warmup', action='store_true',
                       help='Warm up mmap files.')
    group.add_argument('--num-workers', type=int, default=2,
                       help="Dataloader number of workers.")
Mohammad's avatar
Mohammad committed
328
329
330
331
332
    group.add_argument('--tokenizer-type', type=str,
                       default=None,
                       choices=['BertWordPieceLowerCase',
                                'GPT2BPETokenizer'],
                       help='What type of tokenizer to use.')
333
334
335
336
337
338
339
340
341
342
    group.add_argument('--data-impl', type=str, default='infer',
                       choices=['lazy', 'cached', 'mmap', 'infer'],
                       help='Implementation of indexed datasets.')
    group.add_argument('--reset-position-ids', action='store_true',
                       help='Reset posistion ids after end-of-document token.')
    group.add_argument('--reset-attention-mask', action='store_true',
                       help='Reset self attention maske after '
                       'end-of-document token.')
    group.add_argument('--eod-mask-loss', action='store_true',
                       help='Mask loss for the end of document tokens.')
Mohammad's avatar
Mohammad committed
343

Mohammad's avatar
Mohammad committed
344
345
    return parser

Raul Puri's avatar
Raul Puri committed
346

Mohammad's avatar
Mohammad committed
347
348
def _add_autoresume_args(parser):
    group = parser.add_argument_group(title='autoresume')
Raul Puri's avatar
Raul Puri committed
349

Mohammad's avatar
Mohammad committed
350
351
352
353
354
    group.add_argument('--adlr-autoresume', action='store_true',
                       help='Enable autoresume on adlr cluster.')
    group.add_argument('--adlr-autoresume-interval', type=int, default=1000,
                       help='Intervals over which check for autoresume'
                       'termination signal')
Raul Puri's avatar
Raul Puri committed
355

Mohammad's avatar
Mohammad committed
356
    return parser