arguments.py 37.3 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

import argparse
import os

21
import torch
Raul Puri's avatar
Raul Puri committed
22

23
24
def parse_args(extra_args_provider=None, defaults={},
               ignore_unknown_args=False):
Mohammad's avatar
Mohammad committed
25
    """Parse all arguments."""
26
27
    parser = argparse.ArgumentParser(description='Megatron-LM Arguments',
                                     allow_abbrev=False)
Mohammad's avatar
Mohammad committed
28

Mohammad's avatar
Mohammad committed
29
30
31
32
33
34
35
36
37
38
39
40
    # 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)
Mostofa Patwary's avatar
Mostofa Patwary committed
41
    parser = _add_biencoder_args(parser)
42
    parser = _add_vit_args(parser)
43
    parser = _add_logging_args(parser)
Mohammad's avatar
Mohammad committed
44
45
46
47

    # Custom arguments.
    if extra_args_provider is not None:
        parser = extra_args_provider(parser)
Mohammad's avatar
Mohammad committed
48

Mohammad's avatar
Mohammad committed
49
    # Parse.
50
51
52
53
    if ignore_unknown_args:
        args, _ = parser.parse_known_args()
    else:
        args = parser.parse_args()
Mohammad's avatar
Mohammad committed
54

Mohammad's avatar
Mohammad committed
55
56
57
    # Distributed args.
    args.rank = int(os.getenv('RANK', '0'))
    args.world_size = int(os.getenv("WORLD_SIZE", '1'))
mohammad's avatar
mohammad committed
58
    # Tensor model parallel size.
59
60
    args.tensor_model_parallel_size = min(
        args.tensor_model_parallel_size, args.world_size)
mohammad's avatar
mohammad committed
61
62
63
64
    assert args.world_size % args.tensor_model_parallel_size == 0, 'world size'\
        ' ({}) is not divisible by tensor model parallel size ({})'.format(
            args.world_size, args.tensor_model_parallel_size)
    # Pipeline model parallel size.
65
66
67
    args.pipeline_model_parallel_size = min(
        args.pipeline_model_parallel_size,
        (args.world_size // args.tensor_model_parallel_size))
mohammad's avatar
mohammad committed
68
    # Checks.
69
70
71
    model_parallel_size = args.pipeline_model_parallel_size * \
                          args.tensor_model_parallel_size
    assert args.world_size % model_parallel_size == 0, 'world size is not'\
72
        ' divisible by tensor parallel size ({}) times pipeline parallel ' \
mohammad's avatar
mohammad committed
73
74
        'size ({})'.format(args.world_size, args.tensor_model_parallel_size,
                           args.pipeline_model_parallel_size)
75
    args.data_parallel_size = args.world_size // model_parallel_size
Mohammad's avatar
Mohammad committed
76
    if args.rank == 0:
mohammad's avatar
mohammad committed
77
78
79
80
81
82
83
        print('using world size: {}, data-parallel-size: {}, '
              'tensor-model-parallel size: {}, '
              'pipeline-model-parallel size: {} '.format(
                  args.world_size, args.data_parallel_size,
                  args.tensor_model_parallel_size,
                  args.pipeline_model_parallel_size), flush=True)

84
85
86
87
88
89
90
91
92
93
94
    # Deprecated arguments
    assert args.batch_size is None, '--batch-size argument is no longer ' \
        'valid, use --micro-batch-size instead'
    del args.batch_size
    assert args.warmup is None, '--warmup argument is no longer valid, use ' \
        '--lr-warmup-fraction instead'
    del args.warmup
    assert args.model_parallel_size is None, '--model-parallel-size is no ' \
        'longer valid, use --tensor-model-parallel-size instead'
    del args.model_parallel_size

Jared Casper's avatar
Jared Casper committed
95
96
97
98
99
100
101
102
103
104
105
106
107
108
    # Set input defaults.
    for key in defaults:
        # 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.
        if getattr(args, key) is not None:
            if args.rank == 0:
                print('WARNING: overriding default arguments for {key}:{v} \
                       with {key}:{v2}'.format(key=key, v=defaults[key],
                                               v2=getattr(args, key)),
                                               flush=True)
        else:
            setattr(args, key, defaults[key])

mohammad's avatar
mohammad committed
109
110
111
112
113
114
115
116
117
    # Batch size.
    assert args.micro_batch_size is not None
    assert args.micro_batch_size > 0
    if args.global_batch_size is None:
        args.global_batch_size = args.micro_batch_size * args.data_parallel_size
        if args.rank == 0:
            print('setting global batch size to {}'.format(
                args.global_batch_size), flush=True)
    assert args.global_batch_size > 0
118
    if args.num_layers_per_virtual_pipeline_stage is not None:
119
120
121
        assert args.pipeline_model_parallel_size > 2, \
            'pipeline-model-parallel size should be greater than 2 with ' \
            'interleaved schedule'
122
123
124
125
126
127
128
129
        assert args.num_layers % args.num_layers_per_virtual_pipeline_stage == 0, \
            'number of layers is not divisible by number of layers per virtual ' \
            'pipeline stage'
        args.virtual_pipeline_model_parallel_size = \
            (args.num_layers // args.pipeline_model_parallel_size) // \
            args.num_layers_per_virtual_pipeline_stage
    else:
        args.virtual_pipeline_model_parallel_size = None
Mohammad's avatar
Mohammad committed
130

131
132
133
    # Parameters dtype.
    args.params_dtype = torch.float
    if args.fp16:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
134
        assert not args.bf16
135
        args.params_dtype = torch.half
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
136
137
138
    if args.bf16:
        assert not args.fp16
        args.params_dtype = torch.bfloat16
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
139
140
141
142
143
144
145
        # bfloat16 requires gradient accumulation and all-reduce to
        # be done in fp32.
        if not args.accumulate_allreduce_grads_in_fp32:
            args.accumulate_allreduce_grads_in_fp32 = True
            if args.rank == 0:
                print('accumulate and all-reduce gradients in fp32 for '
                      'bfloat16 data type.', flush=True)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
146

147
148
149
150
    if args.rank == 0:
        print('using {} for parameters ...'.format(args.params_dtype),
              flush=True)

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
151
    # If we do accumulation and all-reduces in fp32, we need to have
152
    # local DDP and we should set the use-contiguous-buffers-in-ddp.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
153
154
155
    if args.accumulate_allreduce_grads_in_fp32:
        assert args.DDP_impl == 'local'
        args.use_contiguous_buffers_in_ddp = True
156

157
158
159
    if args.dataloader_type is None:
        args.dataloader_type = 'single'

160
161
162
    # Consumed tokens.
    args.consumed_train_samples = 0
    args.consumed_valid_samples = 0
163

164
165
166
167
168
169
170
171
172
    # Iteration-based training.
    if args.train_iters:
        # If we use iteration-based training, make sure the
        # sample-based options are off.
        assert args.train_samples is None, \
            'expected iteration-based training'
        assert args.lr_decay_samples is None, \
            'expected iteration-based learning rate decay'
        assert args.lr_warmup_samples == 0, \
173
            'expected iteration-based learning rate warmup'
174
175
        assert args.rampup_batch_size is None, \
            'expected no batch-size rampup for iteration-based training'
176
        if args.lr_warmup_fraction is not None:
177
            assert args.lr_warmup_iters == 0, \
178
                'can only specify one of lr-warmup-fraction and lr-warmup-iters'
179
180
181
182
183
184
185
186
187
188
189

    # Sample-based training.
    if args.train_samples:
        # If we use sample-based training, make sure the
        # iteration-based options are off.
        assert args.train_iters is None, \
            'expected sample-based training'
        assert args.lr_decay_iters is None, \
            'expected sample-based learning rate decay'
        assert args.lr_warmup_iters == 0, \
            'expected sample-based learnig rate warmup'
190
        if args.lr_warmup_fraction is not None:
191
            assert args.lr_warmup_samples == 0, \
192
193
                'can only specify one of lr-warmup-fraction ' \
                'and lr-warmup-samples'
194

195
    # Check required arguments.
Mohammad's avatar
Mohammad committed
196
197
    required_args = ['num_layers', 'hidden_size', 'num_attention_heads',
                     'max_position_embeddings']
198
    for req_arg in required_args:
Mohammad's avatar
Mohammad committed
199
        _check_arg_is_not_none(args, req_arg)
200

Mohammad's avatar
Mohammad committed
201
    # Checks.
202
203
204
205
206
207
208
209
210
211
212
213
214
    if args.ffn_hidden_size is None:
        args.ffn_hidden_size = 4 * args.hidden_size

    if args.kv_channels is None:
        assert args.hidden_size % args.num_attention_heads == 0
        args.kv_channels = args.hidden_size // args.num_attention_heads

    if args.seq_length is not None:
        assert args.encoder_seq_length is None
        args.encoder_seq_length = args.seq_length
    else:
        assert args.encoder_seq_length is not None
        args.seq_length = args.encoder_seq_length
215

Mohammad's avatar
Mohammad committed
216
    assert args.hidden_size % args.num_attention_heads == 0
Mohammad's avatar
Mohammad committed
217
218
219
220
    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
221
222
    if args.save is not None:
        assert args.save_interval is not None
mohammad's avatar
mohammad committed
223
224
225
    # Mixed precision checks.
    if args.fp16_lm_cross_entropy:
        assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
226
    if args.fp32_residual_connection:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
227
228
        assert args.fp16 or args.bf16, \
            'residual connection in fp32 only supported when using fp16 or bf16.'
mohammad's avatar
mohammad committed
229
230
231
232
233
    # Activation checkpointing.
    if args.distribute_checkpointed_activations:
        assert args.checkpoint_activations, \
            'for distribute-checkpointed-activations to work you '\
            'need to enable checkpoint-activations'
Vijay Korthikanti's avatar
Vijay Korthikanti committed
234

Mohammad's avatar
Mohammad committed
235
236
    _print_args(args)
    return args
Mohammad's avatar
Mohammad committed
237
238


Mohammad's avatar
Mohammad committed
239
240
241
def _print_args(args):
    """Print arguments."""
    if args.rank == 0:
mohammad's avatar
mohammad committed
242
243
        print('------------------------ arguments ------------------------',
              flush=True)
Mohammad's avatar
Mohammad committed
244
245
        str_list = []
        for arg in vars(args):
mohammad's avatar
mohammad committed
246
            dots = '.' * (48 - len(arg))
Mohammad's avatar
Mohammad committed
247
248
249
            str_list.append('  {} {} {}'.format(arg, dots, getattr(args, arg)))
        for arg in sorted(str_list, key=lambda x: x.lower()):
            print(arg, flush=True)
mohammad's avatar
mohammad committed
250
251
        print('-------------------- end of arguments ---------------------',
              flush=True)
Mohammad's avatar
Mohammad committed
252
253


254
255
256
257
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
258
def _add_network_size_args(parser):
Mohammad's avatar
Mohammad committed
259
    group = parser.add_argument_group(title='network size')
Mohammad's avatar
Mohammad committed
260

261
    group.add_argument('--num-layers', type=int, default=None,
Mohammad's avatar
Mohammad committed
262
                       help='Number of transformer layers.')
263
    group.add_argument('--hidden-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
264
                       help='Tansformer hidden size.')
265
    group.add_argument('--ffn-hidden-size', type=int, default=None,
266
267
                       help='Transformer Feed-Forward Network hidden size. '
                       'This is set to 4*hidden-size if not provided')
268
    group.add_argument('--num-attention-heads', type=int, default=None,
Mohammad's avatar
Mohammad committed
269
                       help='Number of transformer attention heads.')
270
    group.add_argument('--kv-channels', type=int, default=None,
271
272
273
274
                       help='Projection weights dimension in multi-head '
                       'attention. This is set to '
                       '   args.hidden_size // args.num_attention_heads '
                       'if not provided.')
275
    group.add_argument('--max-position-embeddings', type=int, default=None,
Mohammad's avatar
Mohammad committed
276
277
278
279
280
                       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
281
282
    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
                       help='Layer norm epsilon.')
Mohammad's avatar
Mohammad committed
283
284
285
286
    group.add_argument('--apply-residual-connection-post-layernorm',
                       action='store_true',
                       help='If set, use original BERT residula connection '
                       'ordering.')
287
288
289
290
    group.add_argument('--openai-gelu', action='store_true',
                       help='Use OpenAIs GeLU implementation. This option'
                       'should not be used unless for backward compatibility'
                       'reasons.')
291
    group.add_argument('--onnx-safe', type=bool, required=False,
292
293
                       help='Use workarounds for known problems with '
                       'Torch ONNX exporter')
294
295
296
    group.add_argument('--bert-no-binary-head', action='store_false',
                       help='Disable BERT binary head.',
                       dest='bert_binary_head')
Mohammad's avatar
Mohammad committed
297

Mohammad's avatar
Mohammad committed
298
299
300
    return parser


301
302
303
304
305
def _add_logging_args(parser):
    group = parser.add_argument_group(title='logging')

    group.add_argument('--log-params-norm', action='store_true',
                       help='If set, calculate and log parameters norm.')
306
    group.add_argument('--log-num-zeros-in-grad', action='store_true',
Rewon Child's avatar
Rewon Child committed
307
                       help='If set, calculate and log the number of zeros in gradient.')
308
309
    group.add_argument('--tensorboard-log-interval', type=int, default=1,
                       help='Report to tensorboard interval.')
310
311
312
313
    group.add_argument('--tensorboard-queue-size', type=int, default=1000,
                       help='Size of the tensorboard queue for pending events '
                       'and summaries before one of the ‘add’ calls forces a '
                       'flush to disk.')
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
    group.add_argument('--log-timers-to-tensorboard', action='store_true',
                       help='If set, write timers to tensorboard.')
    group.add_argument('--log-batch-size-to-tensorboard', action='store_true',
                       help='If set, write batch-size to tensorboard.')
    group.add_argument('--no-log-learnig-rate-to-tensorboard',
                       action='store_false',
                       help='Disable learning rate logging to tensorboard.',
                       dest='log_learning_rate_to_tensorboard')
    group.add_argument('--no-log-loss-scale-to-tensorboard',
                       action='store_false',
                       help='Disable loss-scale logging to tensorboard.',
                       dest='log_loss_scale_to_tensorboard')
    group.add_argument('--log-validation-ppl-to-tensorboard',
                       action='store_true',
                       help='If set, write validation perplexity to '
                       'tensorboard.')
330
331
332
333

    return parser


Mohammad's avatar
Mohammad committed
334
def _add_regularization_args(parser):
Mohammad's avatar
Mohammad committed
335
336
337
    group = parser.add_argument_group(title='regularization')

    group.add_argument('--attention-dropout', type=float, default=0.1,
338
                       help='Post attention dropout probability.')
Mohammad's avatar
Mohammad committed
339
340
341
342
343
344
    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.')
345
    group.add_argument('--adam-beta1', type=float, default=0.9,
346
347
                       help='First coefficient for computing running averages '
                       'of gradient and its square')
348
    group.add_argument('--adam-beta2', type=float, default=0.999,
349
350
                       help='Second coefficient for computing running averages '
                       'of gradient and its square')
351
    group.add_argument('--adam-eps', type=float, default=1e-08,
352
                       help='Term added to the denominator to improve'
353
                       'numerical stability')
354
355
    group.add_argument('--sgd-momentum', type=float, default=0.9,
                       help='Momentum factor for sgd')
Mohammad's avatar
Mohammad committed
356
357
358

    return parser

Mohammad's avatar
Mohammad committed
359
360

def _add_training_args(parser):
Mohammad's avatar
Mohammad committed
361
362
    group = parser.add_argument_group(title='training')

363
    group.add_argument('--micro-batch-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
364
365
                       help='Batch size per model instance (local batch size). '
                       'Global batch size is local batch size times data '
mohammad's avatar
mohammad committed
366
                       'parallel size times number of micro batches.')
367
368
369
    group.add_argument('--batch-size', type=int, default=None,
                       help='Old batch size parameter, do not use. '
                       'Use --micro-batch-size instead')
mohammad's avatar
mohammad committed
370
    group.add_argument('--global-batch-size', type=int, default=None,
mohammad's avatar
mohammad committed
371
372
373
                       help='Training batch size. If set, it should be a '
                       'multiple of micro-batch-size times data-parallel-size. '
                       'If this value is None, then '
mohammad's avatar
mohammad committed
374
                       'use micro-batch-size * data-parallel-size as the '
mohammad's avatar
mohammad committed
375
376
                       'global batch size. This choice will result in 1 for '
                       'number of micro-batches.')
mohammad's avatar
mohammad committed
377
378
379
380
381
382
383
384
385
386
387
388
    group.add_argument('--rampup-batch-size', nargs='*', default=None,
                       help='Batch size ramp up with the following values:'
                       '  --rampup-batch-size <start batch size> '
                       '                      <batch size incerement> '
                       '                      <ramp-up samples> '
                       'For example:'
                       '   --rampup-batch-size 16 8 300000 \ '
                       '   --global-batch-size 1024'
                       'will start with global batch size 16 and over '
                       ' (1024 - 16) / 8 = 126 intervals will increase'
                       'the batch size linearly to 1024. In each interval'
                       'we will use approximately 300000 / 126 = 2380 samples.')
Mohammad's avatar
Mohammad committed
389
390
391
    group.add_argument('--checkpoint-activations', action='store_true',
                       help='Checkpoint activation to allow for training '
                       'with larger models, sequences, and batch sizes.')
392
393
394
395
    group.add_argument('--distribute-checkpointed-activations',
                       action='store_true',
                       help='If set, distribute checkpointed activations '
                       'across model parallel group.')
Mohammad's avatar
Mohammad committed
396
397
    group.add_argument('--checkpoint-num-layers', type=int, default=1,
                       help='chunk size (number of layers) for checkpointing.')
Mohammad's avatar
Mohammad committed
398
    group.add_argument('--train-iters', type=int, default=None,
Mohammad's avatar
Mohammad committed
399
                       help='Total number of iterations to train over all '
400
401
402
403
404
405
                       'training runs. Note that either train-iters or '
                       'train-samples should be provided.')
    group.add_argument('--train-samples', type=int, default=None,
                       help='Total number of samples to train over all '
                       'training runs. Note that either train-iters or '
                       'train-samples should be provided.')
Mohammad's avatar
Mohammad committed
406
407
408
409
410
    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.')
411
412
    group.add_argument('--exit-duration-in-mins', type=int, default=None,
                       help='Exit the program after this many minutes.')
Mohammad's avatar
Mohammad committed
413
414
    group.add_argument('--tensorboard-dir', type=str, default=None,
                       help='Write TensorBoard logs to this directory.')
415
    group.add_argument('--no-masked-softmax-fusion',
416
417
418
                       action='store_false',
                       help='Disable fusion of query_key_value scaling, '
                       'masking, and softmax.',
419
                       dest='masked_softmax_fusion')
420
421
422
423
424
425
    group.add_argument('--no-bias-gelu-fusion', action='store_false',
                       help='Disable bias and gelu fusion.',
                       dest='bias_gelu_fusion')
    group.add_argument('--no-bias-dropout-fusion', action='store_false',
                       help='Disable bias and dropout fusion.',
                       dest='bias_dropout_fusion')
426
427
428
    group.add_argument('--optimizer', type=str, default='adam',
                       choices=['adam', 'sgd'],
                       help='Optimizer function')
429
    group.add_argument('--dataloader-type', type=str, default=None,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
430
431
                       choices=['single', 'cyclic'],
                       help='Single pass vs multiple pass data loader')
Mohammad's avatar
Mohammad committed
432
433
434
    return parser


Mohammad's avatar
Mohammad committed
435
def _add_initialization_args(parser):
Mohammad's avatar
Mohammad committed
436
437
438
439
440
441
442
443
    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.')
444
445
    group.add_argument('--init-method-xavier-uniform', action='store_true',
                       help='Enable Xavier uniform parameter initialization')
Mohammad's avatar
Mohammad committed
446

Mohammad's avatar
Mohammad committed
447
448
449
    return parser


Mohammad's avatar
Mohammad committed
450
def _add_learning_rate_args(parser):
Mohammad's avatar
Mohammad committed
451
452
    group = parser.add_argument_group(title='learning rate')

Mohammad's avatar
Mohammad committed
453
    group.add_argument('--lr', type=float, default=None,
Mohammad's avatar
Mohammad committed
454
455
456
457
                       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',
mohammad's avatar
mohammad committed
458
                       choices=['constant', 'linear', 'cosine'],
Mohammad's avatar
Mohammad committed
459
460
461
462
                       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`')
463
464
465
    group.add_argument('--lr-decay-samples', type=int, default=None,
                       help='number of samples to decay learning rate over,'
                       ' If None defaults to `--train-samples`')
466
467
468
    group.add_argument('--lr-warmup-fraction', type=float, default=None,
                       help='fraction of lr-warmup-(iters/samples) to use '
                       'for warmup (as a float)')
469
470
471
472
473
474
    group.add_argument('--lr-warmup-iters', type=int, default=0,
                       help='number of iterations to linearly warmup '
                       'learning rate over.')
    group.add_argument('--lr-warmup-samples', type=int, default=0,
                       help='number of samples to linearly warmup '
                       'learning rate over.')
475
    group.add_argument('--warmup', type=int, default=None,
476
                       help='Old lr warmup argument, do not use. Use one of the'
477
                       '--lr-warmup-* arguments above')
Mohammad's avatar
Mohammad committed
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
    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('--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
496
def _add_checkpointing_args(parser):
Mohammad's avatar
Mohammad committed
497
498
499
500
501
502
    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.')
503
    group.add_argument('--no-save-optim', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
504
                       help='Do not save current optimizer.')
505
    group.add_argument('--no-save-rng', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
506
507
508
                       help='Do not save current rng state.')
    group.add_argument('--load', type=str, default=None,
                       help='Directory containing a model checkpoint.')
Jared Casper's avatar
Jared Casper committed
509
    group.add_argument('--no-load-optim', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
510
                       help='Do not load optimizer when loading checkpoint.')
Jared Casper's avatar
Jared Casper committed
511
    group.add_argument('--no-load-rng', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
512
513
514
515
516
517
518
519
520
                       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
521
def _add_mixed_precision_args(parser):
Mohammad's avatar
Mohammad committed
522
523
524
525
    group = parser.add_argument_group(title='mixed precision')

    group.add_argument('--fp16', action='store_true',
                       help='Run model in fp16 mode.')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
526
527
    group.add_argument('--bf16', action='store_true',
                       help='Run model in bfloat16 mode.')
mohammad's avatar
mohammad committed
528
529
530
531
532
533
534
535
536
537
538
539
    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('--initial-loss-scale', type=float, default=2**32,
                       help='Initial loss-scale for dynamic loss scaling.')
    group.add_argument('--min-loss-scale', type=float, default=1.0,
                       help='Minimum loss scale for dynamic loss scale.')
    group.add_argument('--loss-scale-window', type=float, default=1000,
                       help='Window over which to raise/lower dynamic scale.')
    group.add_argument('--hysteresis', type=int, default=2,
                       help='hysteresis for dynamic loss scaling')
540
541
    group.add_argument('--fp32-residual-connection', action='store_true',
                       help='Move residual connections to fp32.')
542
543
544
    group.add_argument('--no-query-key-layer-scaling', action='store_false',
                       help='Do not scale Q * K^T by 1 / layer-number.',
                       dest='apply_query_key_layer_scaling')
Mohammad's avatar
Mohammad committed
545
    group.add_argument('--attention-softmax-in-fp32', action='store_true',
546
547
548
                       help='Run attention masking and softmax in fp32. '
                       'This flag is ignored unless '
                       '--no-query-key-layer-scaling is specified.')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
549
550
551
    group.add_argument('--accumulate-allreduce-grads-in-fp32',
                       action='store_true',
                       help='Gradient accumulation and all-reduce in fp32.')
552
553
554
555
    group.add_argument('--fp16-lm-cross-entropy', action='store_true',
                       help='Move the cross entropy unreduced loss calculation'
                       'for lm head to fp16.')

Mohammad's avatar
Mohammad committed
556
557
558
    return parser


Mohammad's avatar
Mohammad committed
559
def _add_distributed_args(parser):
560
561
    group = parser.add_argument_group(title='distributed')

562
563
564
565
    group.add_argument('--tensor-model-parallel-size', type=int, default=1,
                       help='Degree of tensor model parallelism.')
    group.add_argument('--pipeline-model-parallel-size', type=int, default=1,
                       help='Degree of pipeline model parallelism.')
566
567
568
    group.add_argument('--model-parallel-size', type=int, default=None,
                       help='Old model parallel argument, do not use. Use '
                       '--tensor-model-parallel-size instead.')
569
570
    group.add_argument('--num-layers-per-virtual-pipeline-stage', type=int, default=None,
                       help='Number of layers per virtual pipeline stage')
Mohammad's avatar
Mohammad committed
571
572
573
574
    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
575
                       choices=['local', 'torch'],
Mohammad's avatar
Mohammad committed
576
577
                       help='which DistributedDataParallel implementation '
                       'to use.')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
578
579
580
    group.add_argument('--use-contiguous-buffers-in-ddp', action='store_true',
                       help='If set, use contiguous buffer in DDP. Note that '
                       'this option only works woth local DDP.' )
581
582
583
    group.add_argument('--no-scatter-gather-tensors-in-pipeline', action='store_false',
                       help='Use scatter/gather to optimize communication of tensors in pipeline',
                       dest='scatter_gather_tensors_in_pipeline')
Mohammad's avatar
Mohammad committed
584
585
    group.add_argument('--local_rank', type=int, default=None,
                       help='local rank passed from distributed launcher.')
586
    group.add_argument('--lazy-mpu-init', type=bool, required=False,
587
588
589
590
591
592
593
594
                       help='If set to True, initialize_megatron() '
                       'skips DDP initialization and returns function to '
                       'complete it instead.Also turns on '
                       '--use-cpu-initialization flag. This is for '
                       'external DDP manager.' )
    group.add_argument('--use-cpu-initialization', action='store_true',
                       default=None, help='If set, affine parallel weights '
                       'initialization uses CPU' )
Mohammad's avatar
Mohammad committed
595
596
597
    return parser


Mohammad's avatar
Mohammad committed
598
def _add_validation_args(parser):
Mohammad's avatar
Mohammad committed
599
600
601
602
603
604
605
606
607
    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
608
609
610
    return parser


Mohammad's avatar
Mohammad committed
611
def _add_data_args(parser):
Mohammad's avatar
Mohammad committed
612
613
    group = parser.add_argument_group(title='data and dataloader')

mohammad's avatar
mohammad committed
614
    group.add_argument('--data-path', nargs='*', default=None,
mohammad's avatar
mohammad committed
615
616
617
618
                       help='Path to the training dataset. Accepted format:'
                       '1) a single data path, 2) multiple datasets in the'
                       'form: dataset1-weight dataset1-path dataset2-weight '
                       'dataset2-path ...')
Mohammad's avatar
Mohammad committed
619
    group.add_argument('--split', type=str, default='969, 30, 1',
Mohammad's avatar
Mohammad committed
620
621
                       help='Comma-separated list of proportions for training,'
                       ' validation, and test split. For example the split '
622
623
                       '`90,5,5` will use 90%% of data for training, 5%% for '
                       'validation and 5%% for test.')
Mohammad's avatar
Mohammad committed
624
    group.add_argument('--vocab-file', type=str, default=None,
Mohammad's avatar
Mohammad committed
625
                       help='Path to the vocab file.')
Mohammad's avatar
Mohammad committed
626
627
    group.add_argument('--merge-file', type=str, default=None,
                       help='Path to the BPE merge file.')
628
629
630
    group.add_argument('--vocab-extra-ids', type=int, default=0,
                       help='Number of additional vocabulary tokens. '
                            'They are used for span masking in the T5 model')
Mohammad's avatar
Mohammad committed
631
    group.add_argument('--seq-length', type=int, default=None,
632
                       help='Maximum sequence length to process.')
633
    group.add_argument('--encoder-seq-length', type=int, default=None,
634
635
                       help='Maximum encoder sequence length to process.'
                       'This should be exclusive of --seq-length')
636
637
    group.add_argument('--decoder-seq-length', type=int, default=None,
                       help="Maximum decoder sequence length to process.")
Mostofa Patwary's avatar
Mostofa Patwary committed
638
639
    group.add_argument('--retriever-seq-length', type=int, default=256,
                       help='Maximum sequence length for the biencoder model '
Mostofa Patwary's avatar
Mostofa Patwary committed
640
                        ' for retriever')
641
642
643
    group.add_argument('--sample-rate', type=float, default=1.0,
                       help='sample rate for training data. Supposed to be 0 '
                            ' < sample_rate < 1')
Mohammad's avatar
Mohammad committed
644
645
646
647
648
649
650
651
    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
652
653
654
    group.add_argument('--tokenizer-type', type=str,
                       default=None,
                       choices=['BertWordPieceLowerCase',
Raul Puri's avatar
Raul Puri committed
655
                                'BertWordPieceCase',
Mohammad's avatar
Mohammad committed
656
657
                                'GPT2BPETokenizer'],
                       help='What type of tokenizer to use.')
658
659
660
661
662
663
664
665
666
667
    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
668

Mohammad's avatar
Mohammad committed
669
670
    return parser

Raul Puri's avatar
Raul Puri committed
671

Mohammad's avatar
Mohammad committed
672
673
def _add_autoresume_args(parser):
    group = parser.add_argument_group(title='autoresume')
Raul Puri's avatar
Raul Puri committed
674

Mohammad's avatar
Mohammad committed
675
676
677
678
679
    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
680

Mohammad's avatar
Mohammad committed
681
    return parser
Neel Kant's avatar
Neel Kant committed
682
683


Mostofa Patwary's avatar
Mostofa Patwary committed
684
685
def _add_biencoder_args(parser):
    group = parser.add_argument_group(title='biencoder')
Neel Kant's avatar
Neel Kant committed
686
687
688

    # network size
    group.add_argument('--ict-head-size', type=int, default=None,
689
                       help='Size of block embeddings to be used in ICT and '
Mostofa Patwary's avatar
Mostofa Patwary committed
690
                        'REALM (paper default: 128)')
691
    group.add_argument('--biencoder-projection-dim', type=int, default=0,
Mostofa Patwary's avatar
Mostofa Patwary committed
692
693
                       help='Size of projection head used in biencoder (paper'
                        ' default: 128)')
694
    group.add_argument('--biencoder-shared-query-context-model', action='store_true',
Mostofa Patwary's avatar
Mostofa Patwary committed
695
696
                        help='Whether to share the parameters of the query '
                        'and context models or not')
Neel Kant's avatar
Neel Kant committed
697
698
699
700
701

    # checkpointing
    group.add_argument('--ict-load', type=str, default=None,
                       help='Directory containing an ICTBertModel checkpoint')
    group.add_argument('--bert-load', type=str, default=None,
702
703
                       help='Directory containing an BertModel checkpoint '
                       '(needed to start ICT and REALM)')
Neel Kant's avatar
Neel Kant committed
704
705
706
707
708

    # data
    group.add_argument('--titles-data-path', type=str, default=None,
                       help='Path to titles dataset used for ICT')
    group.add_argument('--query-in-block-prob', type=float, default=0.1,
709
710
                       help='Probability of keeping query in block for '
                       'ICT dataset')
Neel Kant's avatar
Neel Kant committed
711
    group.add_argument('--use-one-sent-docs', action='store_true',
Neel Kant's avatar
Neel Kant committed
712
                       help='Whether to use one sentence documents in ICT')
713
714
    group.add_argument('--evidence-data-path', type=str, default=None,
                       help='Path to Wikipedia Evidence frm DPR paper')
Neel Kant's avatar
Neel Kant committed
715

716
    # training
717
    group.add_argument('--retriever-report-topk-accuracies', nargs='+', type=int,
Mostofa Patwary's avatar
Mostofa Patwary committed
718
719
                        default=[], help="Which top-k accuracies to report "
                        "(e.g. '1 5 20')")
Mostofa Patwary's avatar
Mostofa Patwary committed
720
    group.add_argument('--retriever-score-scaling', action='store_true',
Mostofa Patwary's avatar
Mostofa Patwary committed
721
722
                       help='Whether to scale retriever scores by inverse '
                        'square root of hidden size')
723

Neel Kant's avatar
Neel Kant committed
724
    # faiss index
Neel Kant's avatar
Neel Kant committed
725
    group.add_argument('--block-data-path', type=str, default=None,
Neel Kant's avatar
Neel Kant committed
726
                       help='Where to save/load BlockData to/from')
Mostofa Patwary's avatar
Mostofa Patwary committed
727
728
729
    group.add_argument('--embedding-path', type=str, default=None,
                       help='Where to save/load Open-Retrieval Embedding'
                        ' data to/from')
Neel Kant's avatar
Neel Kant committed
730
731
732

    # indexer
    group.add_argument('--indexer-batch-size', type=int, default=128,
733
734
                       help='How large of batches to use when doing indexing '
                       'jobs')
Neel Kant's avatar
Neel Kant committed
735
    group.add_argument('--indexer-log-interval', type=int, default=1000,
736
737
                       help='After how many batches should the indexer '
                       'report progress')
Neel Kant's avatar
Neel Kant committed
738
    return parser
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753


def _add_vit_args(parser):
    group = parser.add_argument_group(title="vit")

    group.add_argument('--num-classes', type=int, default=1000,
                       help='num of classes in vision classificaiton task')
    group.add_argument('--img-dim', type=int, default=224,
                       help='Image size for vision classification task')
    group.add_argument('--num-channels', type=int, default=3,
                       help='Number of channels in input image data')
    group.add_argument('--patch-dim', type=int, default=16,
                       help='patch dimension used in vit')

    return parser