"docs/backend/sampling_params.md" did not exist on "c9064e6fd9a5356ee579e9d452bfad725f8e6f2c"
arguments.py 19.2 KB
Newer Older
Raul Puri's avatar
Raul Puri committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# coding=utf-8
# Copyright (c) 2019, 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.

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
48
    # Set input defaults.
    for key in defaults:
        setattr(args, key, defaults[key])
Mohammad's avatar
Mohammad committed
49

Mohammad's avatar
Mohammad committed
50
51
52
53
54
55
56
    # 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
57

Mohammad's avatar
Mohammad committed
58
59
60
61
    # Fp16 loss scaling.
    args.dynamic_loss_scale = False
    if args.loss_scale is None:
        args.dynamic_loss_scale = True
Mohammad's avatar
Mohammad committed
62

Mohammad's avatar
Mohammad committed
63

Mohammad's avatar
Mohammad committed
64
65
66
67
68
69
    # Checks.
    assert args.hidden_size % args.num_attention_heads == 0
    assert args.max_position_embeddings >= args.seq_length
    assert args.min_lr <= args.lr
    if args.save is not None:
        assert args.save_interval is not None
Mohammad's avatar
Mohammad committed
70

Mohammad's avatar
Mohammad committed
71
72
    _print_args(args)
    return args
Mohammad's avatar
Mohammad committed
73
74


Mohammad's avatar
Mohammad committed
75
76
77
78
79
80
81
82
83
84
85
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
86
87


Mohammad's avatar
Mohammad committed
88
def _add_network_size_args(parser):
Mohammad's avatar
Mohammad committed
89
    group = parser.add_argument_group(title='network size')
Mohammad's avatar
Mohammad committed
90

Mohammad's avatar
Mohammad committed
91
92
93
94
95
96
97
98
99
100
101
102
    group.add_argument('--num-layers', type=int, required=True,
                       help='Number of transformer layers.')
    group.add_argument('--hidden-size', type=int, required=True,
                       help='Tansformer hidden size.')
    group.add_argument('--num-attention-heads', type=int, required=True,
                       help='Number of transformer attention heads.')
    group.add_argument('--max-position-embeddings', type=int, required=True,
                       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
103
104
105
    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
                       help='Layer norm epsilon.')

Mohammad's avatar
Mohammad committed
106
107
108
    return parser


Mohammad's avatar
Mohammad committed
109
def _add_regularization_args(parser):
Mohammad's avatar
Mohammad committed
110
111
112
113
114
115
116
117
118
119
120
121
122
    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
123
124

def _add_training_args(parser):
Mohammad's avatar
Mohammad committed
125
126
127
128
129
130
131
132
133
134
135
    group = parser.add_argument_group(title='training')

    group.add_argument('--batch-size', type=int, required=True,
                       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
136
    group.add_argument('--train-iters', type=int, default=None,
Mohammad's avatar
Mohammad committed
137
138
139
140
141
142
143
144
145
146
147
148
149
                       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
150
def _add_initialization_args(parser):
Mohammad's avatar
Mohammad committed
151
152
153
154
155
156
157
158
    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
159

Mohammad's avatar
Mohammad committed
160
161
162
    return parser


Mohammad's avatar
Mohammad committed
163
def _add_learning_rate_args(parser):
Mohammad's avatar
Mohammad committed
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
    group = parser.add_argument_group(title='learning rate')

    group.add_argument('--lr', type=float, required=True,
                       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
197
def _add_checkpointing_args(parser):
Mohammad's avatar
Mohammad committed
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
    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
222
def _add_mixed_precision_args(parser):
Mohammad's avatar
Mohammad committed
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
    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.')
    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
247
def _add_distributed_args(parser):
Mohammad's avatar
Mohammad committed
248
249
    group = parser.add_argument_group(title='mixed precision')

Mohammad's avatar
Mohammad committed
250
251
    group.add_argument('--model-parallel-size', type=int, default=1,
                       help='Size of the model parallel.')
Mohammad's avatar
Mohammad committed
252
253
254
255
    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
256
                       choices=['local', 'torch'],
Mohammad's avatar
Mohammad committed
257
258
259
260
261
262
263
264
                       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
265
def _add_validation_args(parser):
Mohammad's avatar
Mohammad committed
266
267
268
269
270
271
272
273
274
    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
275
276
277
    return parser


Mohammad's avatar
Mohammad committed
278
def _add_data_args(parser):
Mohammad's avatar
Mohammad committed
279
280
    group = parser.add_argument_group(title='data and dataloader')

Mohammad's avatar
Mohammad committed
281
    group.add_argument('--data-path', type=str, default=None,
Mohammad's avatar
Mohammad committed
282
                       help='Path to combined dataset to split.')
Mohammad's avatar
Mohammad committed
283
    group.add_argument('--split', type=str, default='969, 30, 1',
Mohammad's avatar
Mohammad committed
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
                       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.')
    group.add_argument('--vocab-file', type=str, required=True,
                       help='Path to the vocab file.')
    group.add_argument('--seq-length', type=int, required=True,
                       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
300
301
302
303
304
305
306
307
    group.add_argument('--tokenizer-type', type=str,
                       default=None,
                       choices=['BertWordPieceLowerCase',
                                'GPT2BPETokenizer'],
                       help='What type of tokenizer to use.')
    parser.add_argument('--data-impl', type=str, default='infer',
                        choices=['lazy', 'cached', 'mmap', 'infer'],
                        help='Implementation of indexed datasets.')
Mohammad's avatar
Mohammad committed
308

Mohammad's avatar
Mohammad committed
309
310
    return parser

Raul Puri's avatar
Raul Puri committed
311

Mohammad's avatar
Mohammad committed
312
313
def _add_autoresume_args(parser):
    group = parser.add_argument_group(title='autoresume')
Raul Puri's avatar
Raul Puri committed
314

Mohammad's avatar
Mohammad committed
315
316
317
318
319
    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
320

Mohammad's avatar
Mohammad committed
321
    return parser
Raul Puri's avatar
Raul Puri committed
322
323


Mohammad's avatar
Mohammad committed
324
########################################################################
Raul Puri's avatar
Raul Puri committed
325
326


Mohammad's avatar
Mohammad committed
327
def add_training_args_(parser):
Raul Puri's avatar
Raul Puri committed
328
329
330
331
    """Training arguments."""

    group = parser.add_argument_group('train', 'training configurations')

332
333
334
335
336
337
    # Batch prodecuer arguments
    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.')
338
339
    group.add_argument('--eod-mask-loss', action='store_true',
                       help='Mask loss for the end of document tokens')
340

Raul Puri's avatar
Raul Puri committed
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
    return parser


def add_evaluation_args(parser):
    """Evaluation arguments."""

    group = parser.add_argument_group('validation', 'validation configurations')

    group.add_argument('--eval-batch-size', type=int, default=None,
                       help='Data Loader batch size for evaluation datasets.'
                       'Defaults to `--batch-size`')
    group.add_argument('--eval-seq-length', type=int, default=None,
                       help='Maximum sequence length to process for '
                       'evaluation. Defaults to `--seq-length`')
    group.add_argument('--eval-max-preds-per-seq', type=int, default=None,
                       help='Maximum number of predictions to use for '
                       'evaluation. Defaults to '
                       'math.ceil(`--eval-seq-length`*.15/10)*10')
359
360
361
362
    group.add_argument('--overlapping-eval', type=int, default=32,
                       help='sliding window for overlapping eval ')
    group.add_argument('--cloze-eval', action='store_true',
                       help='Evaluation dataset from `--valid-data` is a cloze task')
363
364
    group.add_argument('--strict-lambada', action='store_true',
                       help='use more difficult formulation of lambada')
365
366
367
368
369
370
    group.add_argument('--eval-hf', action='store_true',
                       help='perform evaluation with huggingface openai model.'
                       'use `--load` to specify weights path to be loaded')
    group.add_argument('--load-openai', action='store_true',
                       help='load openai weights into our model. Use `--load` '
                       'to specify weights path to be loaded')
Raul Puri's avatar
Raul Puri committed
371
372
373

    return parser

374
375
376
377
378
def add_text_generate_args(parser):
    """Text generate arguments."""

    group = parser.add_argument_group('Text generation', 'configurations')
    group.add_argument("--temperature", type=float, default=1.0)
379
    group.add_argument("--greedy", action='store_true', default=False)
380
381
    group.add_argument("--top_p", type=float, default=0.0)
    group.add_argument("--top_k", type=int, default=0)
382
383
384
385
386
387
388
389
390
391
392
393
394
395
    group.add_argument("--out-seq-length", type=int, default=1024)
    group.add_argument("--sample-input-file", type=str, default="",
                      help='get input from file instead of interactive mode, '
                           'each line is an input' )
    group.add_argument("--sample-output-file", type=str, default="",
                      help='output file got from --sample-input-file')
    group.add_argument("--num-samples", type=int, default=0,
                       help='number of samples to generate unconditionally, '
                       'defaults to 0 and interactive conditional sampling')
    group.add_argument("--genfile", type=str,
                       help='output file when generating unconditionally')
    group.add_argument("--recompute", action='store_true',
                       help='during generation recompute all attention '
                       'instead of using previously computed keys/values.')
396
397
    return parser

Raul Puri's avatar
Raul Puri committed
398

Mohammad's avatar
Mohammad committed
399
def add_data_args_(parser):
Raul Puri's avatar
Raul Puri committed
400
401
402
403
    """Train/valid/test data arguments."""

    group = parser.add_argument_group('data', 'data configurations')

404
405
406
407
    group.add_argument('--data-loader', type=str, default=None,
                       choices=['raw', 'lazy', 'tfrecords', 'numpy', 'binary'],
                       help='Which data loader to use. Default varies by model.')

408
    group.add_argument('--train-data', nargs='+', default=None,
409
                       help='Whitespace separated paths or corpora names '
Raul Puri's avatar
Raul Puri committed
410
                       'for training.')
411
412
413
414
    group.add_argument('--valid-data', nargs='*', default=None,
                       help='path(s) to the validation data.')
    group.add_argument('--test-data', nargs='*', default=None,
                       help='path(s) to the testing data.')
Mohammad's avatar
Mohammad committed
415

416
417
    # arguments for binary data loader
    # arguments for numpy data loader
418
    group.add_argument('--input-data-sizes-file', type=str, default='sizes.txt',
419
                       help='the filename containing all the shards sizes for numpy data loader')
420

Raul Puri's avatar
Raul Puri committed
421
422
    return parser