arguments.py 30.3 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
22
from megatron import fused_kernels
Raul Puri's avatar
Raul Puri committed
23

24
25
def parse_args(extra_args_provider=None, defaults={},
               ignore_unknown_args=False):
Mohammad's avatar
Mohammad committed
26
    """Parse all arguments."""
27
28
    parser = argparse.ArgumentParser(description='Megatron-LM Arguments',
                                     allow_abbrev=False)
Mohammad's avatar
Mohammad committed
29

Mohammad's avatar
Mohammad committed
30
31
32
33
34
35
36
37
38
39
40
41
    # Standard arguments.
    parser = _add_network_size_args(parser)
    parser = _add_regularization_args(parser)
    parser = _add_training_args(parser)
    parser = _add_initialization_args(parser)
    parser = _add_learning_rate_args(parser)
    parser = _add_checkpointing_args(parser)
    parser = _add_mixed_precision_args(parser)
    parser = _add_distributed_args(parser)
    parser = _add_validation_args(parser)
    parser = _add_data_args(parser)
    parser = _add_autoresume_args(parser)
Neel Kant's avatar
Neel Kant committed
42
    parser = _add_realm_args(parser)
Mohammad's avatar
Mohammad committed
43
44
45
46

    # Custom arguments.
    if extra_args_provider is not None:
        parser = extra_args_provider(parser)
Mohammad's avatar
Mohammad committed
47

Mohammad's avatar
Mohammad committed
48
    # Parse.
49
50
51
52
    if ignore_unknown_args:
        args, _ = parser.parse_known_args()
    else:
        args = parser.parse_args()
Mohammad's avatar
Mohammad committed
53

Mohammad's avatar
Mohammad committed
54
55
56
    # Distributed args.
    args.rank = int(os.getenv('RANK', '0'))
    args.world_size = int(os.getenv("WORLD_SIZE", '1'))
mohammad's avatar
mohammad committed
57
    # Tensor model parallel size.
58
59
    args.tensor_model_parallel_size = min(
        args.tensor_model_parallel_size, args.world_size)
mohammad's avatar
mohammad committed
60
61
62
63
    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.
64
65
66
    args.pipeline_model_parallel_size = min(
        args.pipeline_model_parallel_size,
        (args.world_size // args.tensor_model_parallel_size))
mohammad's avatar
mohammad committed
67
    # Checks.
68
69
70
    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'\
mohammad's avatar
mohammad committed
71
72
73
        ' divisible by tensor parallel size ({}) times pipeline paralle ' \
        'size ({})'.format(args.world_size, args.tensor_model_parallel_size,
                           args.pipeline_model_parallel_size)
74
    args.data_parallel_size = args.world_size // model_parallel_size
Mohammad's avatar
Mohammad committed
75
    if args.rank == 0:
mohammad's avatar
mohammad committed
76
77
78
79
80
81
82
        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)

83
84
85
86
87
88
89
90
91
92
93
    # 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

mohammad's avatar
mohammad committed
94
95
96
97
98
99
100
101
102
    # 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
Mohammad's avatar
Mohammad committed
103

104
105
106
107
108
109
110
111
    # Parameters dtype.
    args.params_dtype = torch.float
    if args.fp16:
        args.params_dtype = torch.half
    if args.rank == 0:
        print('using {} for parameters ...'.format(args.params_dtype),
              flush=True)

112
113
114
    # Consumed tokens.
    args.consumed_train_samples = 0
    args.consumed_valid_samples = 0
115

Mohammad's avatar
Mohammad committed
116
117
    # Set input defaults.
    for key in defaults:
Mohammad's avatar
Mohammad committed
118
119
120
        # 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.
Raul Puri's avatar
Raul Puri committed
121
        if getattr(args, key) is not None:
Raul Puri's avatar
Raul Puri committed
122
            if args.rank == 0:
Raul Puri's avatar
Raul Puri committed
123
124
                print('WARNING: overriding default arguments for {key}:{v} \
                       with {key}:{v2}'.format(key=key, v=defaults[key],
Raul Puri's avatar
Raul Puri committed
125
126
                                               v2=getattr(args, key)),
                                               flush=True)
Raul Puri's avatar
Raul Puri committed
127
128
        else:
            setattr(args, key, defaults[key])
Mohammad's avatar
Mohammad committed
129

130
131
132
133
134
135
136
137
138
    # 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, \
139
            'expected iteration-based learning rate warmup'
140
141
        assert args.rampup_batch_size is None, \
            'expected no batch-size rampup for iteration-based training'
142
        if args.lr_warmup_fraction is not None:
143
            assert args.lr_warmup_iters == 0, \
144
                'can only specify one of lr-warmup-fraction and lr-warmup-iters'
145
146
147
148
149
150
151
152
153
154
155

    # 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'
156
        if args.lr_warmup_fraction is not None:
157
            assert args.lr_warmup_samples == 0, \
158
                'can only specify one of lr-warmup-fraction and lr-warmup-samples'
159

160
    # Check required arguments.
Mohammad's avatar
Mohammad committed
161
162
    required_args = ['num_layers', 'hidden_size', 'num_attention_heads',
                     'max_position_embeddings']
163
    for req_arg in required_args:
Mohammad's avatar
Mohammad committed
164
        _check_arg_is_not_none(args, req_arg)
165

Mohammad's avatar
Mohammad committed
166
    # Checks.
167
168
169
170
171
172
173
174
175
176
177
178
179
180
    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
 
Mohammad's avatar
Mohammad committed
181
    assert args.hidden_size % args.num_attention_heads == 0
Mohammad's avatar
Mohammad committed
182
183
184
185
    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
186
187
    if args.save is not None:
        assert args.save_interval is not None
mohammad's avatar
mohammad committed
188
189
190
    # Mixed precision checks.
    if args.fp16_lm_cross_entropy:
        assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
191
192
    if args.fp32_residual_connection:
        assert args.fp16, \
mshoeybi's avatar
mshoeybi committed
193
            'residual connection in fp32 only supported when using fp16.'
mohammad's avatar
mohammad committed
194
195
196
197
198
    # Activation checkpointing.
    if args.distribute_checkpointed_activations:
        assert args.checkpoint_activations, \
            'for distribute-checkpointed-activations to work you '\
            'need to enable checkpoint-activations'
199
200
201
202
203
   
    # Load scaled_masked_softmax_fusion_kernels
    if args.masked_softmax_fusion:
        fused_kernels.load_scaled_upper_triang_masked_softmax_fusion_kernel()
        fused_kernels.load_scaled_masked_softmax_fusion_kernel()
204

205
206
207
208
    # Load mixed precision fused layer norm.
    if args.fp32_residual_connection:
        fused_kernels.load_fused_mix_prec_layer_norm_kernel()

Mohammad's avatar
Mohammad committed
209
210
    _print_args(args)
    return args
Mohammad's avatar
Mohammad committed
211
212


Mohammad's avatar
Mohammad committed
213
214
215
def _print_args(args):
    """Print arguments."""
    if args.rank == 0:
mohammad's avatar
mohammad committed
216
217
        print('------------------------ arguments ------------------------',
              flush=True)
Mohammad's avatar
Mohammad committed
218
219
        str_list = []
        for arg in vars(args):
mohammad's avatar
mohammad committed
220
            dots = '.' * (48 - len(arg))
Mohammad's avatar
Mohammad committed
221
222
223
            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
224
225
        print('-------------------- end of arguments ---------------------',
              flush=True)
Mohammad's avatar
Mohammad committed
226
227


228
229
230
231
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
232
def _add_network_size_args(parser):
Mohammad's avatar
Mohammad committed
233
    group = parser.add_argument_group(title='network size')
Mohammad's avatar
Mohammad committed
234

235
    group.add_argument('--num-layers', type=int, default=None,
Mohammad's avatar
Mohammad committed
236
                       help='Number of transformer layers.')
237
    group.add_argument('--hidden-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
238
                       help='Tansformer hidden size.')
239
240
241
    group.add_argument('--ffn-hidden-size', type=int, default=None,
                       help='Transformer Feed-Forward Network hidden size. This is set to 4*hidden-size if not '
                            'provided')
242
    group.add_argument('--num-attention-heads', type=int, default=None,
Mohammad's avatar
Mohammad committed
243
                       help='Number of transformer attention heads.')
244
245
246
    group.add_argument('--kv-channels', type=int, default=None,
                       help='Projection weights dimension in multi-head attention. '
                            'This is set to args.hidden_size // args.num_attention_heads if not provided.')
247
    group.add_argument('--max-position-embeddings', type=int, default=None,
Mohammad's avatar
Mohammad committed
248
249
250
251
252
                       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
253
254
    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
                       help='Layer norm epsilon.')
Mohammad's avatar
Mohammad committed
255
256
257
258
    group.add_argument('--apply-residual-connection-post-layernorm',
                       action='store_true',
                       help='If set, use original BERT residula connection '
                       'ordering.')
259
260
261
262
    group.add_argument('--openai-gelu', action='store_true',
                       help='Use OpenAIs GeLU implementation. This option'
                       'should not be used unless for backward compatibility'
                       'reasons.')
263
    group.add_argument('--onnx-safe', type=bool, required=False,
264
                       help='Use workarounds for known problems with Torch ONNX exporter')
Mohammad's avatar
Mohammad committed
265

Mohammad's avatar
Mohammad committed
266
267
268
    return parser


Mohammad's avatar
Mohammad committed
269
def _add_regularization_args(parser):
Mohammad's avatar
Mohammad committed
270
271
272
    group = parser.add_argument_group(title='regularization')

    group.add_argument('--attention-dropout', type=float, default=0.1,
273
                       help='Post attention dropout probability.')
Mohammad's avatar
Mohammad committed
274
275
276
277
278
279
    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.')
280
281
282
283
284
285
286
    group.add_argument('--adam-beta1', type=float, default=0.9,
                       help='First coefficient for computing running averages of'
                       'gradient and its square')
    group.add_argument('--adam-beta2', type=float, default=0.999,
                       help='Second coefficient for computing running averages of'
                       'gradient and its square')
    group.add_argument('--adam-eps', type=float, default=1e-08,
287
                       help='Term added to the denominator to improve'
288
                       'numerical stability')
Mohammad's avatar
Mohammad committed
289
290
291

    return parser

Mohammad's avatar
Mohammad committed
292
293

def _add_training_args(parser):
Mohammad's avatar
Mohammad committed
294
295
    group = parser.add_argument_group(title='training')

296
    group.add_argument('--micro-batch-size', type=int, default=None,
Mohammad's avatar
Mohammad committed
297
298
                       help='Batch size per model instance (local batch size). '
                       'Global batch size is local batch size times data '
mohammad's avatar
mohammad committed
299
                       'parallel size times number of micro batches.')
300
301
302
    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
303
    group.add_argument('--global-batch-size', type=int, default=None,
mohammad's avatar
mohammad committed
304
305
306
                       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
307
                       'use micro-batch-size * data-parallel-size as the '
mohammad's avatar
mohammad committed
308
309
                       'global batch size. This choice will result in 1 for '
                       'number of micro-batches.')
mohammad's avatar
mohammad committed
310
311
312
313
314
315
316
317
318
319
320
321
    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
322
323
324
    group.add_argument('--checkpoint-activations', action='store_true',
                       help='Checkpoint activation to allow for training '
                       'with larger models, sequences, and batch sizes.')
325
326
327
328
    group.add_argument('--distribute-checkpointed-activations',
                       action='store_true',
                       help='If set, distribute checkpointed activations '
                       'across model parallel group.')
Mohammad's avatar
Mohammad committed
329
330
    group.add_argument('--checkpoint-num-layers', type=int, default=1,
                       help='chunk size (number of layers) for checkpointing.')
Mohammad's avatar
Mohammad committed
331
    group.add_argument('--train-iters', type=int, default=None,
Mohammad's avatar
Mohammad committed
332
                       help='Total number of iterations to train over all '
333
334
335
336
337
338
                       '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
339
340
341
342
343
    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.')
344
345
    group.add_argument('--exit-duration-in-mins', type=int, default=None,
                       help='Exit the program after this many minutes.')
Mohammad's avatar
Mohammad committed
346
347
    group.add_argument('--tensorboard-dir', type=str, default=None,
                       help='Write TensorBoard logs to this directory.')
348
    group.add_argument('--no-masked-softmax-fusion',
349
350
351
                       action='store_false',
                       help='Disable fusion of query_key_value scaling, '
                       'masking, and softmax.',
352
                       dest='masked_softmax_fusion')
353
354
355
356
357
358
    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')
Mohammad's avatar
Mohammad committed
359
360
361
362

    return parser


Mohammad's avatar
Mohammad committed
363
def _add_initialization_args(parser):
Mohammad's avatar
Mohammad committed
364
365
366
367
368
369
370
371
    group = parser.add_argument_group(title='initialization')

    group.add_argument('--seed', type=int, default=1234,
                       help='Random seed used for python, numpy, '
                       'pytorch, and cuda.')
    group.add_argument('--init-method-std', type=float, default=0.02,
                       help='Standard deviation of the zero mean normal '
                       'distribution used for weight initialization.')
Mohammad's avatar
Mohammad committed
372

Mohammad's avatar
Mohammad committed
373
374
375
    return parser


Mohammad's avatar
Mohammad committed
376
def _add_learning_rate_args(parser):
Mohammad's avatar
Mohammad committed
377
378
    group = parser.add_argument_group(title='learning rate')

Mohammad's avatar
Mohammad committed
379
    group.add_argument('--lr', type=float, default=None,
Mohammad's avatar
Mohammad committed
380
381
382
383
                       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
384
                       choices=['constant', 'linear', 'cosine'],
Mohammad's avatar
Mohammad committed
385
386
387
388
                       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`')
389
390
391
    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`')
392
393
394
    group.add_argument('--lr-warmup-fraction', type=float, default=None,
                       help='fraction of lr-warmup-(iters/samples) to use '
                       'for warmup (as a float)')
395
396
397
398
399
400
    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.')
401
402
403
    group.add_argument('--warmup', type=int, default=None,
                       help='Old lr warmup argument, do not use. Use one of the '
                       '--lr-warmup-* arguments above')
Mohammad's avatar
Mohammad committed
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
    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
422
def _add_checkpointing_args(parser):
Mohammad's avatar
Mohammad committed
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
    group = parser.add_argument_group(title='checkpointing')

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

    return parser


Mohammad's avatar
Mohammad committed
447
def _add_mixed_precision_args(parser):
Mohammad's avatar
Mohammad committed
448
449
450
451
    group = parser.add_argument_group(title='mixed precision')

    group.add_argument('--fp16', action='store_true',
                       help='Run model in fp16 mode.')
mohammad's avatar
mohammad committed
452
453
454
455
456
457
458
459
460
461
462
463
    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')
464
465
    group.add_argument('--fp32-residual-connection', action='store_true',
                       help='Move residual connections to fp32.')
466
467
468
    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
469
    group.add_argument('--attention-softmax-in-fp32', action='store_true',
470
471
472
                       help='Run attention masking and softmax in fp32. '
                       'This flag is ignored unless '
                       '--no-query-key-layer-scaling is specified.')
Mohammad's avatar
Mohammad committed
473
474
    group.add_argument('--fp32-allreduce', action='store_true',
                       help='All-reduce in fp32')
475
476
477
478
    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
479
480
481
    return parser


Mohammad's avatar
Mohammad committed
482
def _add_distributed_args(parser):
483
484
    group = parser.add_argument_group(title='distributed')

485
486
487
488
    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.')
489
490
491
    group.add_argument('--model-parallel-size', type=int, default=None,
                       help='Old model parallel argument, do not use. Use '
                       '--tensor-model-parallel-size instead.')
Mohammad's avatar
Mohammad committed
492
493
494
495
    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
496
                       choices=['local', 'torch'],
Mohammad's avatar
Mohammad committed
497
498
499
500
                       help='which DistributedDataParallel implementation '
                       'to use.')
    group.add_argument('--local_rank', type=int, default=None,
                       help='local rank passed from distributed launcher.')
501
502
    group.add_argument('--lazy-mpu-init', type=bool, required=False,
                       help='If set to True, initialize_megatron() skips DDP initialization'
Boris Fomitchev's avatar
Boris Fomitchev committed
503
504
                       ' and returns function to complete it instead.'
                       'Also turns on --use-cpu-initialization flag.'
505
                       'This is for external DDP manager.' )
506
507
    group.add_argument('--use-cpu-initialization', action='store_true',
                       help='If set, affine parallel weights initialization uses CPU' )
Mohammad's avatar
Mohammad committed
508
509
510
    return parser


Mohammad's avatar
Mohammad committed
511
def _add_validation_args(parser):
Mohammad's avatar
Mohammad committed
512
513
514
515
516
517
518
519
520
    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
521
522
523
    return parser


Mohammad's avatar
Mohammad committed
524
def _add_data_args(parser):
Mohammad's avatar
Mohammad committed
525
526
    group = parser.add_argument_group(title='data and dataloader')

mohammad's avatar
mohammad committed
527
    group.add_argument('--data-path', nargs='*', default=None,
mohammad's avatar
mohammad committed
528
529
530
531
                       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
532
    group.add_argument('--split', type=str, default='969, 30, 1',
Mohammad's avatar
Mohammad committed
533
534
                       help='Comma-separated list of proportions for training,'
                       ' validation, and test split. For example the split '
535
536
                       '`90,5,5` will use 90%% of data for training, 5%% for '
                       'validation and 5%% for test.')
Mohammad's avatar
Mohammad committed
537
    group.add_argument('--vocab-file', type=str, default=None,
Mohammad's avatar
Mohammad committed
538
                       help='Path to the vocab file.')
Mohammad's avatar
Mohammad committed
539
540
    group.add_argument('--merge-file', type=str, default=None,
                       help='Path to the BPE merge file.')
Mohammad's avatar
Mohammad committed
541
    group.add_argument('--seq-length', type=int, default=None,
Mohammad's avatar
Mohammad committed
542
                       help="Maximum sequence length to process.")
543
544
545
546
    group.add_argument('--encoder-seq-length', type=int, default=None,
                       help="Maximum encoder sequence length to process.")
    group.add_argument('--decoder-seq-length', type=int, default=None,
                       help="Maximum decoder sequence length to process.")
Mohammad's avatar
Mohammad committed
547
548
549
550
551
552
553
554
    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
555
556
557
    group.add_argument('--tokenizer-type', type=str,
                       default=None,
                       choices=['BertWordPieceLowerCase',
Raul Puri's avatar
Raul Puri committed
558
                                'BertWordPieceCase',
Mohammad's avatar
Mohammad committed
559
560
                                'GPT2BPETokenizer'],
                       help='What type of tokenizer to use.')
561
562
563
564
565
566
567
568
569
570
    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
571

Mohammad's avatar
Mohammad committed
572
573
    return parser

Raul Puri's avatar
Raul Puri committed
574

Mohammad's avatar
Mohammad committed
575
576
def _add_autoresume_args(parser):
    group = parser.add_argument_group(title='autoresume')
Raul Puri's avatar
Raul Puri committed
577

Mohammad's avatar
Mohammad committed
578
579
580
581
582
    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
583

Mohammad's avatar
Mohammad committed
584
    return parser
Neel Kant's avatar
Neel Kant committed
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604


def _add_realm_args(parser):
    group = parser.add_argument_group(title='realm')

    # network size
    group.add_argument('--ict-head-size', type=int, default=None,
                       help='Size of block embeddings to be used in ICT and REALM (paper default: 128)')

    # 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,
                       help='Directory containing an BertModel checkpoint (needed to start ICT and REALM)')

    # 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,
                       help='Probability of keeping query in block for ICT dataset')
Neel Kant's avatar
Neel Kant committed
605
    group.add_argument('--use-one-sent-docs', action='store_true',
Neel Kant's avatar
Neel Kant committed
606
607
                       help='Whether to use one sentence documents in ICT')

608
609
610
611
    # training
    group.add_argument('--report-topk-accuracies', nargs='+', default=[],
                       help="Which top-k accuracies to report (e.g. '1 5 20')")

Neel Kant's avatar
Neel Kant committed
612
613
614
    # faiss index
    group.add_argument('--faiss-use-gpu', action='store_true',
                       help='Whether create the FaissMIPSIndex on GPU')
Neel Kant's avatar
Neel Kant committed
615
    group.add_argument('--block-data-path', type=str, default=None,
Neel Kant's avatar
Neel Kant committed
616
                       help='Where to save/load BlockData to/from')
Neel Kant's avatar
Neel Kant committed
617
618
619
620
621
622

    # indexer
    group.add_argument('--indexer-batch-size', type=int, default=128,
                       help='How large of batches to use when doing indexing jobs')
    group.add_argument('--indexer-log-interval', type=int, default=1000,
                       help='After how many batches should the indexer report progress')
Neel Kant's avatar
Neel Kant committed
623
    return parser