arguments.py 41.8 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)
mshoeybi's avatar
mshoeybi committed
44
    parser = _add_inference_args(parser)
Mohammad's avatar
Mohammad committed
45
46
47
48

    # Custom arguments.
    if extra_args_provider is not None:
        parser = extra_args_provider(parser)
Mohammad's avatar
Mohammad committed
49

Mohammad's avatar
Mohammad committed
50
    # Parse.
51
52
53
54
    if ignore_unknown_args:
        args, _ = parser.parse_known_args()
    else:
        args = parser.parse_args()
Mohammad's avatar
Mohammad committed
55

Mohammad's avatar
Mohammad committed
56
57
58
    # Distributed args.
    args.rank = int(os.getenv('RANK', '0'))
    args.world_size = int(os.getenv("WORLD_SIZE", '1'))
mohammad's avatar
mohammad committed
59
    # Tensor model parallel size.
60
61
    args.tensor_model_parallel_size = min(
        args.tensor_model_parallel_size, args.world_size)
mohammad's avatar
mohammad committed
62
63
64
65
    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.
66
67
68
    args.pipeline_model_parallel_size = min(
        args.pipeline_model_parallel_size,
        (args.world_size // args.tensor_model_parallel_size))
mohammad's avatar
mohammad committed
69
    # Checks.
70
71
72
    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'\
73
        ' divisible by tensor parallel size ({}) times pipeline parallel ' \
mohammad's avatar
mohammad committed
74
75
        'size ({})'.format(args.world_size, args.tensor_model_parallel_size,
                           args.pipeline_model_parallel_size)
76
    args.data_parallel_size = args.world_size // model_parallel_size
Mohammad's avatar
Mohammad committed
77
    if args.rank == 0:
mohammad's avatar
mohammad committed
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
    if args.pipeline_model_parallel_size > 1:
        if args.pipeline_model_parallel_split_rank is not None:
            assert args.pipeline_model_parallel_split_rank < \
                    args.pipeline_model_parallel_size, 'split rank needs'\
                    ' to be less than pipeline model parallel size ({})'.format(
                            args.pipeline_model_parallel_size)
mohammad's avatar
mohammad committed
90

91
92
93
94
95
96
97
98
99
100
    # 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
101
102
    if args.checkpoint_activations:
        args.activations_checkpoint_method = 'uniform'
slym's avatar
slym committed
103
104
105
106
        if args.rank == 0:
            print('--checkpoint-activations is no longer valid, '
                  'use --activation-checkpoint-method instead. '
                  'Defaulting to activation-checkpoint-method=uniform.')
107
    del args.checkpoint_activations
108

Jared Casper's avatar
Jared Casper committed
109
110
111
112
113
114
115
116
117
118
119
120
121
122
    # 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
123
124
125
126
127
128
129
130
131
    # 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
132
    if args.num_layers_per_virtual_pipeline_stage is not None:
133
134
135
        assert args.pipeline_model_parallel_size > 2, \
            'pipeline-model-parallel size should be greater than 2 with ' \
            'interleaved schedule'
136
137
138
139
140
141
142
143
        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
144

145
146
147
    # Parameters dtype.
    args.params_dtype = torch.float
    if args.fp16:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
148
        assert not args.bf16
149
        args.params_dtype = torch.half
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
150
151
152
    if args.bf16:
        assert not args.fp16
        args.params_dtype = torch.bfloat16
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
153
154
155
156
157
158
159
        # 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
160

161
162
163
164
    if args.rank == 0:
        print('using {} for parameters ...'.format(args.params_dtype),
              flush=True)

165
166
    # If we do accumulation and all-reduces in fp32, we need to have local DDP
    # and we should make sure use-contiguous-buffers-in-local-ddp is not off.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
167
168
    if args.accumulate_allreduce_grads_in_fp32:
        assert args.DDP_impl == 'local'
169
        assert args.use_contiguous_buffers_in_local_ddp
170

mshoeybi's avatar
mshoeybi committed
171
172
173
174
    # For torch DDP, we do not use contiguous buffer
    if args.DDP_impl == 'torch':
        args.use_contiguous_buffers_in_local_ddp = False

175
176
177
    if args.dataloader_type is None:
        args.dataloader_type = 'single'

178
179
180
    # Consumed tokens.
    args.consumed_train_samples = 0
    args.consumed_valid_samples = 0
181

182
183
184
185
186
187
188
189
190
    # 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, \
191
            'expected iteration-based learning rate warmup'
192
193
        assert args.rampup_batch_size is None, \
            'expected no batch-size rampup for iteration-based training'
194
        if args.lr_warmup_fraction is not None:
195
            assert args.lr_warmup_iters == 0, \
196
                'can only specify one of lr-warmup-fraction and lr-warmup-iters'
197
198
199
200
201
202
203
204
205
206
207

    # 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'
208
        if args.lr_warmup_fraction is not None:
209
            assert args.lr_warmup_samples == 0, \
210
211
                'can only specify one of lr-warmup-fraction ' \
                'and lr-warmup-samples'
212

213
    # Check required arguments.
Mohammad's avatar
Mohammad committed
214
215
    required_args = ['num_layers', 'hidden_size', 'num_attention_heads',
                     'max_position_embeddings']
216
    for req_arg in required_args:
Mohammad's avatar
Mohammad committed
217
        _check_arg_is_not_none(args, req_arg)
218

Mohammad's avatar
Mohammad committed
219
    # Checks.
220
221
222
223
224
225
226
227
228
229
230
231
232
    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
233

Mohammad's avatar
Mohammad committed
234
235
    if args.seq_length is not None:
        assert args.max_position_embeddings >= args.seq_length
Jared Casper's avatar
Jared Casper committed
236
237
    if args.decoder_seq_length is not None:
        assert args.max_position_embeddings >= args.decoder_seq_length
Mohammad's avatar
Mohammad committed
238
239
    if args.lr is not None:
        assert args.min_lr <= args.lr
Mohammad's avatar
Mohammad committed
240
241
    if args.save is not None:
        assert args.save_interval is not None
mohammad's avatar
mohammad committed
242
243
244
    # Mixed precision checks.
    if args.fp16_lm_cross_entropy:
        assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
245
    if args.fp32_residual_connection:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
246
247
        assert args.fp16 or args.bf16, \
            'residual connection in fp32 only supported when using fp16 or bf16.'
mohammad's avatar
mohammad committed
248
249
    # Activation checkpointing.
    if args.distribute_checkpointed_activations:
mshoeybi's avatar
mshoeybi committed
250
251
252
        assert args.tensor_model_parallel_size > 1, 'can distribute ' \
            'checkpointed activations only across tensor model ' \
            'parallel groups'
253
        assert args.activations_checkpoint_method is not None, \
mohammad's avatar
mohammad committed
254
            'for distribute-checkpointed-activations to work you '\
mshoeybi's avatar
mshoeybi committed
255
            'need to use a activation-checkpoint method '
mshoeybi's avatar
tested  
mshoeybi committed
256
257
258
        assert args.num_layers_per_virtual_pipeline_stage is None, \
            'currently distrobuted checkpoint activations only supported for ' \
            'nointerleaved pipeline parallelism'
Vijay Korthikanti's avatar
Vijay Korthikanti committed
259

Sangkug Lym's avatar
Sangkug Lym committed
260
261
262
263
264
265
266
267
268
269
    TORCH_MAJOR = int(torch.__version__.split('.')[0])
    TORCH_MINOR = int(torch.__version__.split('.')[1])
    # Persistent fused layer norm.
    if TORCH_MAJOR < 1 or (TORCH_MAJOR == 1 and TORCH_MINOR < 11):
        args.no_persist_layer_norm = True
        if args.rank == 0:
            print('Persistent fused layer norm kernel is supported from '
                  'pytorch v1.11 (nvidia pytorch container paired with v1.11). '
                  'Defaulting to no_persist_layer_norm=True')

Mohammad's avatar
Mohammad committed
270
271
    _print_args(args)
    return args
Mohammad's avatar
Mohammad committed
272
273


Mohammad's avatar
Mohammad committed
274
275
276
def _print_args(args):
    """Print arguments."""
    if args.rank == 0:
mohammad's avatar
mohammad committed
277
278
        print('------------------------ arguments ------------------------',
              flush=True)
Mohammad's avatar
Mohammad committed
279
280
        str_list = []
        for arg in vars(args):
mohammad's avatar
mohammad committed
281
            dots = '.' * (48 - len(arg))
Mohammad's avatar
Mohammad committed
282
283
284
            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
285
286
        print('-------------------- end of arguments ---------------------',
              flush=True)
Mohammad's avatar
Mohammad committed
287
288


289
290
291
292
def _check_arg_is_not_none(args, arg):
    assert getattr(args, arg) is not None, '{} argument is None'.format(arg)


mshoeybi's avatar
mshoeybi committed
293
294
295
296
297
298
299
300
301
302
303
304
def _add_inference_args(parser):
    group = parser.add_argument_group(title='inference')

    group.add_argument('--inference-batch-times-seqlen-threshold',
                       type=int, default=512,
                       help='During inference, if batch-size times '
                       'sequence-length is smaller than this threshold '
                       'then we will not use pipelining, otherwise we will.')

    return parser

    
Mohammad's avatar
Mohammad committed
305
def _add_network_size_args(parser):
Mohammad's avatar
Mohammad committed
306
    group = parser.add_argument_group(title='network size')
Mohammad's avatar
Mohammad committed
307

308
    group.add_argument('--num-layers', type=int, default=None,
Mohammad's avatar
Mohammad committed
309
                       help='Number of transformer layers.')
310
    group.add_argument('--hidden-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
311
                       help='Tansformer hidden size.')
312
    group.add_argument('--ffn-hidden-size', type=int, default=None,
313
314
                       help='Transformer Feed-Forward Network hidden size. '
                       'This is set to 4*hidden-size if not provided')
315
    group.add_argument('--num-attention-heads', type=int, default=None,
Mohammad's avatar
Mohammad committed
316
                       help='Number of transformer attention heads.')
317
    group.add_argument('--kv-channels', type=int, default=None,
318
319
320
321
                       help='Projection weights dimension in multi-head '
                       'attention. This is set to '
                       '   args.hidden_size // args.num_attention_heads '
                       'if not provided.')
322
    group.add_argument('--max-position-embeddings', type=int, default=None,
Mohammad's avatar
Mohammad committed
323
324
325
326
327
                       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
328
329
    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
                       help='Layer norm epsilon.')
Mohammad's avatar
Mohammad committed
330
331
332
333
    group.add_argument('--apply-residual-connection-post-layernorm',
                       action='store_true',
                       help='If set, use original BERT residula connection '
                       'ordering.')
334
335
336
337
    group.add_argument('--openai-gelu', action='store_true',
                       help='Use OpenAIs GeLU implementation. This option'
                       'should not be used unless for backward compatibility'
                       'reasons.')
338
    group.add_argument('--onnx-safe', type=bool, required=False,
339
340
                       help='Use workarounds for known problems with '
                       'Torch ONNX exporter')
341
342
343
    group.add_argument('--bert-no-binary-head', action='store_false',
                       help='Disable BERT binary head.',
                       dest='bert_binary_head')
Mohammad's avatar
Mohammad committed
344

Mohammad's avatar
Mohammad committed
345
346
347
    return parser


348
349
350
351
352
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.')
353
    group.add_argument('--log-num-zeros-in-grad', action='store_true',
Rewon Child's avatar
Rewon Child committed
354
                       help='If set, calculate and log the number of zeros in gradient.')
355
356
    group.add_argument('--tensorboard-log-interval', type=int, default=1,
                       help='Report to tensorboard interval.')
357
358
359
360
    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.')
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
    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.')
377
378
    group.add_argument('--log-memory-to-tensorboard',
                       action='store_true',
379
                       help='Enable memory logging to tensorboard.')
380
381
382
383

    return parser


Mohammad's avatar
Mohammad committed
384
def _add_regularization_args(parser):
Mohammad's avatar
Mohammad committed
385
386
387
    group = parser.add_argument_group(title='regularization')

    group.add_argument('--attention-dropout', type=float, default=0.1,
388
                       help='Post attention dropout probability.')
Mohammad's avatar
Mohammad committed
389
390
391
392
393
394
    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.')
395
    group.add_argument('--adam-beta1', type=float, default=0.9,
396
397
                       help='First coefficient for computing running averages '
                       'of gradient and its square')
398
    group.add_argument('--adam-beta2', type=float, default=0.999,
399
400
                       help='Second coefficient for computing running averages '
                       'of gradient and its square')
401
    group.add_argument('--adam-eps', type=float, default=1e-08,
402
                       help='Term added to the denominator to improve'
403
                       'numerical stability')
404
405
    group.add_argument('--sgd-momentum', type=float, default=0.9,
                       help='Momentum factor for sgd')
Mohammad's avatar
Mohammad committed
406
407
408

    return parser

Mohammad's avatar
Mohammad committed
409
410

def _add_training_args(parser):
Mohammad's avatar
Mohammad committed
411
412
    group = parser.add_argument_group(title='training')

413
    group.add_argument('--micro-batch-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
414
415
                       help='Batch size per model instance (local batch size). '
                       'Global batch size is local batch size times data '
mohammad's avatar
mohammad committed
416
                       'parallel size times number of micro batches.')
417
418
419
    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
420
    group.add_argument('--global-batch-size', type=int, default=None,
mohammad's avatar
mohammad committed
421
422
423
                       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
424
                       'use micro-batch-size * data-parallel-size as the '
mohammad's avatar
mohammad committed
425
426
                       'global batch size. This choice will result in 1 for '
                       'number of micro-batches.')
mohammad's avatar
mohammad committed
427
428
429
430
431
432
433
434
435
436
437
438
    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
439
440
441
    group.add_argument('--checkpoint-activations', action='store_true',
                       help='Checkpoint activation to allow for training '
                       'with larger models, sequences, and batch sizes.')
442
443
444
445
    group.add_argument('--distribute-checkpointed-activations',
                       action='store_true',
                       help='If set, distribute checkpointed activations '
                       'across model parallel group.')
446
447
448
449
450
    group.add_argument('--activations-checkpoint-method', type=str, default=None,
                       choices=['uniform', 'block'],
                       help='1) uniform: uniformly divide the total number of '
                       'Transformer layers and checkpoint the input activation of '
                       'each divided chunk, '
slym's avatar
slym committed
451
452
453
454
                       '2) checkpoint the input activations of only a set number of '
                       'individual Transformer layers per pipeline stage and do the '
                       'rest without any checkpointing'
                       'default) do not apply activations checkpoint to any layers')
455
456
457
458
459
    group.add_argument('--activations-checkpoint-num-layers', type=int, default=1,
                       help='1) uniform: the number of Transformer layers in each '
                       'uniformly divided checkpoint unit, '
                       '2) block: the number of individual Transformer layers '
                       'to checkpoint within each pipeline stage.')
Mohammad's avatar
Mohammad committed
460
    group.add_argument('--train-iters', type=int, default=None,
Mohammad's avatar
Mohammad committed
461
                       help='Total number of iterations to train over all '
462
463
464
465
466
467
                       '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
468
469
470
471
472
    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.')
473
474
    group.add_argument('--exit-duration-in-mins', type=int, default=None,
                       help='Exit the program after this many minutes.')
Mohammad's avatar
Mohammad committed
475
476
    group.add_argument('--tensorboard-dir', type=str, default=None,
                       help='Write TensorBoard logs to this directory.')
477
    group.add_argument('--no-masked-softmax-fusion',
478
479
480
                       action='store_false',
                       help='Disable fusion of query_key_value scaling, '
                       'masking, and softmax.',
481
                       dest='masked_softmax_fusion')
482
483
484
485
486
487
    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')
488
489
490
    group.add_argument('--optimizer', type=str, default='adam',
                       choices=['adam', 'sgd'],
                       help='Optimizer function')
491
    group.add_argument('--dataloader-type', type=str, default=None,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
492
493
                       choices=['single', 'cyclic'],
                       help='Single pass vs multiple pass data loader')
slym's avatar
slym committed
494
495
496
497
498
    group.add_argument('--no-async-tensor-model-parallel-allreduce',
                       action='store_true',
                       help='Disable asynchronous execution of '
                       'tensor-model-parallel all-reduce with weight '
                       'gradient compuation of a column-linear layer.')
Sangkug Lym's avatar
Sangkug Lym committed
499
500
501
502
503
    group.add_argument('--no-persist-layer-norm', action='store_true',
                       help='Disable using persistent fused layer norm kernel. '
                       'This kernel supports only a set of hidden sizes. Please '
                       'check persist_ln_hidden_sizes if your hidden '
                       'size is supported.')
Mohammad's avatar
Mohammad committed
504
505
506
    return parser


Mohammad's avatar
Mohammad committed
507
def _add_initialization_args(parser):
Mohammad's avatar
Mohammad committed
508
509
510
511
512
513
514
515
    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.')
516
517
    group.add_argument('--init-method-xavier-uniform', action='store_true',
                       help='Enable Xavier uniform parameter initialization')
Mohammad's avatar
Mohammad committed
518

Mohammad's avatar
Mohammad committed
519
520
521
    return parser


Mohammad's avatar
Mohammad committed
522
def _add_learning_rate_args(parser):
Mohammad's avatar
Mohammad committed
523
524
    group = parser.add_argument_group(title='learning rate')

Mohammad's avatar
Mohammad committed
525
    group.add_argument('--lr', type=float, default=None,
Mohammad's avatar
Mohammad committed
526
527
528
529
                       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
530
                       choices=['constant', 'linear', 'cosine'],
Mohammad's avatar
Mohammad committed
531
532
533
534
                       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`')
535
536
537
    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`')
538
539
540
    group.add_argument('--lr-warmup-fraction', type=float, default=None,
                       help='fraction of lr-warmup-(iters/samples) to use '
                       'for warmup (as a float)')
541
542
543
544
545
546
    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.')
547
    group.add_argument('--warmup', type=int, default=None,
548
                       help='Old lr warmup argument, do not use. Use one of the'
549
                       '--lr-warmup-* arguments above')
Mohammad's avatar
Mohammad committed
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
    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
568
def _add_checkpointing_args(parser):
Mohammad's avatar
Mohammad committed
569
570
571
572
573
574
    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.')
575
    group.add_argument('--no-save-optim', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
576
                       help='Do not save current optimizer.')
577
    group.add_argument('--no-save-rng', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
578
579
580
                       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
581
    group.add_argument('--no-load-optim', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
582
                       help='Do not load optimizer when loading checkpoint.')
Jared Casper's avatar
Jared Casper committed
583
    group.add_argument('--no-load-rng', action='store_true', default=None,
Mohammad's avatar
Mohammad committed
584
585
586
587
588
589
590
591
592
                       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
593
def _add_mixed_precision_args(parser):
Mohammad's avatar
Mohammad committed
594
595
596
597
    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
598
599
    group.add_argument('--bf16', action='store_true',
                       help='Run model in bfloat16 mode.')
mohammad's avatar
mohammad committed
600
601
602
603
604
605
606
607
608
609
610
611
    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')
612
613
    group.add_argument('--fp32-residual-connection', action='store_true',
                       help='Move residual connections to fp32.')
614
615
616
    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
617
    group.add_argument('--attention-softmax-in-fp32', action='store_true',
618
619
620
                       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
621
622
623
    group.add_argument('--accumulate-allreduce-grads-in-fp32',
                       action='store_true',
                       help='Gradient accumulation and all-reduce in fp32.')
624
625
626
627
    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
628
629
630
    return parser


Mohammad's avatar
Mohammad committed
631
def _add_distributed_args(parser):
632
633
    group = parser.add_argument_group(title='distributed')

634
635
636
637
    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.')
638
639
640
    group.add_argument('--pipeline-model-parallel-split-rank',
                       type=int, default=None,
                       help='Rank where encoder and decoder should be split.')
641
642
643
    group.add_argument('--model-parallel-size', type=int, default=None,
                       help='Old model parallel argument, do not use. Use '
                       '--tensor-model-parallel-size instead.')
644
645
    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
646
647
648
649
    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
650
                       choices=['local', 'torch'],
Mohammad's avatar
Mohammad committed
651
652
                       help='which DistributedDataParallel implementation '
                       'to use.')
653
654
655
656
    group.add_argument('--no-contiguous-buffers-in-local-ddp',
                       action='store_false', help='If set, dont use '
                       'contiguous buffer in local DDP.',
                       dest='use_contiguous_buffers_in_local_ddp')
657
658
659
    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
660
661
    group.add_argument('--local_rank', type=int, default=None,
                       help='local rank passed from distributed launcher.')
662
    group.add_argument('--lazy-mpu-init', type=bool, required=False,
663
664
665
666
667
668
669
670
                       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' )
Lawrence McAfee's avatar
Lawrence McAfee committed
671
    group.add_argument('--empty-unused-memory-level', default=0, type=int,
672
673
674
675
                       choices=[0, 1, 2],
                       help='Call torch.cuda.empty_cache() each iteration '
                       '(training and eval), to reduce fragmentation.'
                       '0=off, 1=moderate, 2=aggressive.')
Mohammad's avatar
Mohammad committed
676
677
678
    return parser


Mohammad's avatar
Mohammad committed
679
def _add_validation_args(parser):
Mohammad's avatar
Mohammad committed
680
681
682
683
684
685
686
687
688
    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
689
690
691
    return parser


Mohammad's avatar
Mohammad committed
692
def _add_data_args(parser):
Mohammad's avatar
Mohammad committed
693
694
    group = parser.add_argument_group(title='data and dataloader')

mohammad's avatar
mohammad committed
695
    group.add_argument('--data-path', nargs='*', default=None,
mohammad's avatar
mohammad committed
696
697
698
699
                       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
700
    group.add_argument('--split', type=str, default='969, 30, 1',
Mohammad's avatar
Mohammad committed
701
702
                       help='Comma-separated list of proportions for training,'
                       ' validation, and test split. For example the split '
703
704
                       '`90,5,5` will use 90%% of data for training, 5%% for '
                       'validation and 5%% for test.')
Mohammad's avatar
Mohammad committed
705
    group.add_argument('--vocab-file', type=str, default=None,
Mohammad's avatar
Mohammad committed
706
                       help='Path to the vocab file.')
Mohammad's avatar
Mohammad committed
707
708
    group.add_argument('--merge-file', type=str, default=None,
                       help='Path to the BPE merge file.')
709
710
711
    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
712
    group.add_argument('--seq-length', type=int, default=None,
713
                       help='Maximum sequence length to process.')
714
    group.add_argument('--encoder-seq-length', type=int, default=None,
715
716
                       help='Maximum encoder sequence length to process.'
                       'This should be exclusive of --seq-length')
717
718
    group.add_argument('--decoder-seq-length', type=int, default=None,
                       help="Maximum decoder sequence length to process.")
Mostofa Patwary's avatar
Mostofa Patwary committed
719
720
    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
721
                        ' for retriever')
722
723
724
    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
725
726
727
728
729
730
731
732
    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
733
734
735
    group.add_argument('--tokenizer-type', type=str,
                       default=None,
                       choices=['BertWordPieceLowerCase',
Raul Puri's avatar
Raul Puri committed
736
                                'BertWordPieceCase',
Mohammad's avatar
Mohammad committed
737
738
                                'GPT2BPETokenizer'],
                       help='What type of tokenizer to use.')
739
740
741
742
743
744
745
746
747
748
    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
749

Mohammad's avatar
Mohammad committed
750
751
    return parser

Raul Puri's avatar
Raul Puri committed
752

Mohammad's avatar
Mohammad committed
753
754
def _add_autoresume_args(parser):
    group = parser.add_argument_group(title='autoresume')
Raul Puri's avatar
Raul Puri committed
755

Mohammad's avatar
Mohammad committed
756
757
758
759
760
    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
761

Mohammad's avatar
Mohammad committed
762
    return parser
Neel Kant's avatar
Neel Kant committed
763
764


Mostofa Patwary's avatar
Mostofa Patwary committed
765
766
def _add_biencoder_args(parser):
    group = parser.add_argument_group(title='biencoder')
Neel Kant's avatar
Neel Kant committed
767
768
769

    # network size
    group.add_argument('--ict-head-size', type=int, default=None,
770
                       help='Size of block embeddings to be used in ICT and '
Mostofa Patwary's avatar
Mostofa Patwary committed
771
                        'REALM (paper default: 128)')
772
    group.add_argument('--biencoder-projection-dim', type=int, default=0,
Mostofa Patwary's avatar
Mostofa Patwary committed
773
774
                       help='Size of projection head used in biencoder (paper'
                        ' default: 128)')
775
    group.add_argument('--biencoder-shared-query-context-model', action='store_true',
Mostofa Patwary's avatar
Mostofa Patwary committed
776
777
                        help='Whether to share the parameters of the query '
                        'and context models or not')
Neel Kant's avatar
Neel Kant committed
778
779
780
781
782

    # 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,
783
784
                       help='Directory containing an BertModel checkpoint '
                       '(needed to start ICT and REALM)')
Neel Kant's avatar
Neel Kant committed
785
786
787
788
789

    # 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,
790
791
                       help='Probability of keeping query in block for '
                       'ICT dataset')
Neel Kant's avatar
Neel Kant committed
792
    group.add_argument('--use-one-sent-docs', action='store_true',
Neel Kant's avatar
Neel Kant committed
793
                       help='Whether to use one sentence documents in ICT')
794
795
    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
796

797
    # training
798
    group.add_argument('--retriever-report-topk-accuracies', nargs='+', type=int,
Mostofa Patwary's avatar
Mostofa Patwary committed
799
800
                        default=[], help="Which top-k accuracies to report "
                        "(e.g. '1 5 20')")
Mostofa Patwary's avatar
Mostofa Patwary committed
801
    group.add_argument('--retriever-score-scaling', action='store_true',
Mostofa Patwary's avatar
Mostofa Patwary committed
802
803
                       help='Whether to scale retriever scores by inverse '
                        'square root of hidden size')
804

Neel Kant's avatar
Neel Kant committed
805
    # faiss index
Neel Kant's avatar
Neel Kant committed
806
    group.add_argument('--block-data-path', type=str, default=None,
Neel Kant's avatar
Neel Kant committed
807
                       help='Where to save/load BlockData to/from')
Mostofa Patwary's avatar
Mostofa Patwary committed
808
809
810
    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
811
812
813

    # indexer
    group.add_argument('--indexer-batch-size', type=int, default=128,
814
815
                       help='How large of batches to use when doing indexing '
                       'jobs')
Neel Kant's avatar
Neel Kant committed
816
    group.add_argument('--indexer-log-interval', type=int, default=1000,
817
818
                       help='After how many batches should the indexer '
                       'report progress')
Neel Kant's avatar
Neel Kant committed
819
    return parser
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834


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