-
Notifications
You must be signed in to change notification settings - Fork 984
/
arguments.py
570 lines (490 loc) · 28.1 KB
/
arguments.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
# coding=utf-8
#
# Copyright 2021 Biderman et al. This file is based on code by the authors denoted below and has been modified from its original version.
#
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://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.
"""Megatron arguments."""
import argparse
import os
import torch
from megatron import fused_kernels
import deepspeed
def _get_parser(extra_args_provider=None):
parser = argparse.ArgumentParser(description='Megatron-LM Arguments',
allow_abbrev=False)
# 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)
parser = _add_realm_args(parser)
parser = _add_zero_args(parser)
parser = _add_activation_checkpoint_args(parser)
# Custom arguments.
if extra_args_provider is not None:
parser = extra_args_provider(parser)
# Include DeepSpeed configuration arguments
parser = deepspeed.add_config_arguments(parser)
return parser
def parse_args(extra_args_provider=None, defaults={},
ignore_unknown_args=False):
"""Parse all arguments."""
parser = _get_parser(extra_args_provider)
# Parse.
if ignore_unknown_args:
args, _ = parser.parse_known_args()
else:
args = parser.parse_args()
# Distributed args.
args.rank = int(os.getenv('RANK', '0'))
args.world_size = int(os.getenv("WORLD_SIZE", '1'))
args.model_parallel_size = min(args.model_parallel_size, args.world_size)
if args.rank == 0:
print('using world size: {} and model-parallel size: {} '.format(
args.world_size, args.model_parallel_size))
# Fp16 loss scaling.
args.dynamic_loss_scale = False
if args.loss_scale is None:
args.dynamic_loss_scale = True
# 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)
# 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])
# Check required arguments.
required_args = ['num_layers', 'hidden_size', 'num_attention_heads',
'max_position_embeddings']
for req_arg in required_args:
_check_arg_is_not_none(args, req_arg)
# Checks.
assert args.hidden_size % args.num_attention_heads == 0
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
if args.save is not None:
assert args.save_interval is not None
# Parameters sharing does not work with torch DDP.
if (args.num_unique_layers is not None) and (args.num_layers is not None):
assert args.num_unique_layers <= args.num_layers
assert args.num_layers % args.num_unique_layers == 0, \
'num-layers should be divisible by num-unique-layers.'
if args.num_unique_layers < args.num_layers:
assert args.DDP_impl == 'local', \
'torch-DDP does not work with parameters sharing.'
# Mixed precision checks.
if args.fp16_lm_cross_entropy:
assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
# Activation checkpointing.
if args.distribute_checkpointed_activations:
assert args.checkpoint_activations, \
'for distribute-checkpointed-activations to work you ' \
'need to enable checkpoint-activations'
# load scaled_upper_triang_masked_softmax_fusion kernel
if args.scaled_upper_triang_masked_softmax_fusion:
fused_kernels.load_scaled_upper_triang_masked_softmax_fusion_kernel()
# load scaled_masked_softmax_fusion kernel
if args.scaled_masked_softmax_fusion:
fused_kernels.load_scaled_masked_softmax_fusion_kernel()
_print_args(args)
return args
def _print_args(args):
"""Print arguments."""
if args.rank == 0:
print('-------------------- arguments --------------------', flush=True)
str_list = []
for arg in vars(args):
dots = '.' * (32 - len(arg))
str_list.append(' {} {} {}'.format(arg, dots, getattr(args, arg)))
for arg in sorted(str_list, key=lambda x: x.lower()):
print(arg, flush=True)
print('---------------- end of arguments ----------------', flush=True)
def _check_arg_is_not_none(args, arg):
assert getattr(args, arg) is not None, '{} argument is None'.format(arg)
def _add_network_size_args(parser):
group = parser.add_argument_group(title='network size')
group.add_argument('--num-layers', type=int, default=None,
help='Number of transformer layers.')
group.add_argument('--num-unique-layers', type=int, default=None,
help='Number of unique transformer layers. '
'`num-layers` should be divisible by this value.')
group.add_argument('--param-sharing-style', default='grouped',
choices=['grouped', 'spaced'],
help='Ordering of the shared parameters. For example, '
'for a `num-layers`=4 and `--num-unique-layers`=2, '
'we will have the following ordering for two unique '
'layers 1 and 2: '
' grouped: [1, 2, 1, 2] and spaced: [1, 1, 2, 2].')
group.add_argument('--hidden-size', type=int, default=None,
help='Tansformer hidden size.')
group.add_argument('--num-attention-heads', type=int, default=None,
help='Number of transformer attention heads.')
group.add_argument('--max-position-embeddings', type=int, default=None,
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.')
group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
help='Layer norm epsilon.')
group.add_argument('--apply-residual-connection-post-layernorm',
action='store_true',
help='If set, use original BERT residula connection '
'ordering.')
group.add_argument('--openai-gelu', action='store_true',
help='Use OpenAIs GeLU implementation. This option'
'should not be used unless for backward compatibility'
'reasons.')
group.add_argument('--onnx-safe', type=bool, required=False,
help='Use workarounds for known problems with Torch ONNX exporter')
return parser
def _add_regularization_args(parser):
group = parser.add_argument_group(title='regularization')
group.add_argument('--attention-dropout', type=float, default=0.1,
help='Post attention dropout probability.')
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.')
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,
help='Term added to the denominator to improve'
'numerical stability')
return parser
def _add_training_args(parser):
group = parser.add_argument_group(title='training')
group.add_argument('--batch-size', type=int, default=None,
help='Batch size per model instance (local batch size). '
'Global batch size is local batch size times data '
'parallel size.')
group.add_argument('--onebitadam', action='store_true',
help='Enable one bit adam optimizer [MUST BE USING DEEPSPEED]')
group.add_argument('--gas', type=int, default=1,
help='Gradient accumulation steps (pipeline parallelism only). '
'Global batch size is local batch size times data '
'parallel size times gas.')
group.add_argument('--checkpoint-activations', action='store_true',
help='Checkpoint activation to allow for training '
'with larger models, sequences, and batch sizes.')
group.add_argument('--distribute-checkpointed-activations',
action='store_true',
help='If set, distribute checkpointed activations '
'across model parallel group.')
group.add_argument('--checkpoint-num-layers', type=int, default=1,
help='chunk size (number of layers) for checkpointing.')
group.add_argument('--train-iters', type=int, default=None,
help='Total number of iterations to train over all '
'training runs.')
group.add_argument('--log-interval', type=int, default=100,
help='Report loss and timing interval.')
group.add_argument('--exit-interval', type=int, default=None,
help='Exit the program after the iteration is divisible '
'by this value.')
group.add_argument('--tensorboard-dir', type=str, default=None,
help='Write TensorBoard logs to this directory.')
group.add_argument('--scaled-upper-triang-masked-softmax-fusion',
action='store_true',
help='Enable fusion of query_key_value_scaling '
'time (upper diagonal) masking and softmax.')
group.add_argument('--scaled-masked-softmax-fusion',
action='store_true',
help='Enable fusion of query_key_value_scaling '
'general masking and softmax.')
group.add_argument('--bias-gelu-fusion', action='store_true',
help='Enable bias and gelu fusion.')
group.add_argument('--geglu', action='store_true',
help='Enable geglu activation function (WARNING: will increase memory usage, '
'adjust embd dims accordingly)')
group.add_argument('--no-weight-tying', action='store_true',
help='Disables weight tying between embedding weights and final Linear layer')
group.add_argument('--sinusoidal-pos-emb', action='store_true',
help='Uses Sinusoidal Positional embedding applied to the inputs instead of learned')
group.add_argument('--rpe', action='store_true',
help='T5 relative positional encoding')
group.add_argument('--rpe-causal', action='store_true',
help='T5 relative positional encoding causal flag')
group.add_argument('--rpe-num-buckets', type=int, default=32,
help='T5 relative positional encoding number of buckets, default 32.')
group.add_argument('--rpe-max-distance', type=int, default=128,
help='T5 relative positional encoding max distance, default 128.')
group.add_argument('--bias-dropout-fusion', action='store_true',
help='Enable bias and dropout fusion.')
group.add_argument('--sparsity', type=str, default='none',
choices=['none', 'all', 'interspersed'],
help='sparse attention layer configuration. \
none = all regular attn, \
all = all sparse attn, \
interspersed = sparse on odd layers, dense on even')
group.add_argument('--rms-norm', action='store_true',
help='Use root mean squared norm')
group.add_argument('--rms-norm-epsilon', type=float, default=1e-8)
group.add_argument('--cpu-optimizer', action='store_true',
help='Run optimizer on CPU')
group.add_argument('--cpu_torch_adam', action='store_true',
help='Use Torch Adam as optimizer on CPU.')
return parser
def _add_initialization_args(parser):
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.')
return parser
def _add_learning_rate_args(parser):
group = parser.add_argument_group(title='learning rate')
group.add_argument('--lr', type=float, default=None,
help='Initial learning rate. Depending on decay style '
'and initial warmup, the learing rate at each '
'iteration would be different.')
group.add_argument('--lr-decay-style', type=str, default='linear',
choices=['constant', 'linear', 'cosine', 'exponential'],
help='Learning rate decay function.')
group.add_argument('--lr-decay-iters', type=int, default=None,
help='number of iterations to decay learning rate over,'
' If None defaults to `--train-iters`')
group.add_argument('--min-lr', type=float, default=0.0,
help='Minumum value for learning rate. The scheduler'
'clip values below this threshold.')
group.add_argument('--warmup', type=float, default=0.01,
help='Percentage of total iterations to warmup on '
'(.01 = 1 percent of all training iters).')
group.add_argument('--override-lr-scheduler', action='store_true',
help='Reset the values of the scheduler (learning rate,'
'warmup iterations, minimum learning rate, maximum '
'number of iterations, and decay style from input '
'arguments and ignore values from checkpoints. Note'
'that all the above values will be reset.')
group.add_argument('--use-checkpoint-lr-scheduler', action='store_true',
help='Use checkpoint to set the values of the scheduler '
'(learning rate, warmup iterations, minimum learning '
'rate, maximum number of iterations, and decay style '
'from checkpoint and ignore input arguments.')
return parser
def _add_checkpointing_args(parser):
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
def _add_mixed_precision_args(parser):
group = parser.add_argument_group(title='mixed precision')
group.add_argument('--fp16', action='store_true',
help='Run model in fp16 mode.')
group.add_argument('--apply-query-key-layer-scaling', action='store_true',
help='Scale Q * K^T by 1 / layer-number. If this flag '
'is set, then it will automatically set '
'attention-softmax-in-fp32 to true')
group.add_argument('--attention-softmax-in-fp32', action='store_true',
help='Run attention masking and softmax in fp32.')
group.add_argument('--fp32-allreduce', action='store_true',
help='All-reduce in fp32')
group.add_argument('--hysteresis', type=int, default=2,
help='hysteresis for dynamic loss scaling')
group.add_argument('--loss-scale', type=float, default=None,
help='Static loss scaling, positive power of 2 '
'values can improve fp16 convergence. If None, dynamic'
'loss scaling is used.')
group.add_argument('--loss-scale-window', type=float, default=1000,
help='Window over which to raise/lower dynamic scale.')
group.add_argument('--min-scale', type=float, default=1,
help='Minimum loss scale for dynamic loss scale.')
group.add_argument('--fp16-lm-cross-entropy', action='store_true',
help='Move the cross entropy unreduced loss calculation'
'for lm head to fp16.')
return parser
def _add_distributed_args(parser):
group = parser.add_argument_group(title='mixed precision')
group.add_argument('--model-parallel-size', type=int, default=1,
help='Size of the model parallel.')
group.add_argument('--pipe-parallel-size', type=int, default=0,
help='Size of the pipeline parallel. Disable with 0.')
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',
choices=['local', 'torch'],
help='which DistributedDataParallel implementation '
'to use.')
group.add_argument('--local_rank', type=int, default=None,
help='local rank passed from distributed launcher.')
group.add_argument('--lazy-mpu-init', type=bool, required=False,
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',
help='If set, affine parallel weights initialization uses CPU')
return parser
def _add_validation_args(parser):
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.')
return parser
def _add_data_args(parser):
group = parser.add_argument_group(title='data and dataloader')
group.add_argument('--data-path', type=str, default=None,
help='Path to combined dataset to split.')
group.add_argument('--split', type=str, default='969, 30, 1',
help='Comma-separated list of proportions for training,'
' validation, and test split. For example the split '
'`90,5,5` will use 90% of data for training, 5% for '
'validation and 5% for test.')
group.add_argument('--vocab-file', type=str, default=None,
help='Path to the vocab file.')
group.add_argument('--merge-file', type=str, default=None,
help='Path to the BPE merge file.')
group.add_argument('--seq-length', type=int, default=None,
help="Maximum sequence length to process.")
group.add_argument('--mask-prob', type=float, default=0.15,
help='Probability of replacing a token with mask.')
group.add_argument('--short-seq-prob', type=float, default=0.1,
help='Probability of producing a short sequence.')
group.add_argument('--mmap-warmup', action='store_true',
help='Warm up mmap files.')
group.add_argument('--num-workers', type=int, default=2,
help="Dataloader number of workers.")
group.add_argument('--tokenizer-type', type=str,
default=None,
choices=['BertWordPieceLowerCase',
'BertWordPieceCase',
'GPT2BPETokenizer'],
help='What type of tokenizer to use.')
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.')
return parser
def _add_autoresume_args(parser):
group = parser.add_argument_group(title='autoresume')
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')
return parser
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')
group.add_argument('--use-one-sent-docs', action='store_true',
help='Whether to use one sentence documents in ICT')
# training
group.add_argument('--report-topk-accuracies', nargs='+', default=[],
help="Which top-k accuracies to report (e.g. '1 5 20')")
# faiss index
group.add_argument('--faiss-use-gpu', action='store_true',
help='Whether create the FaissMIPSIndex on GPU')
group.add_argument('--block-data-path', type=str, default=None,
help='Where to save/load BlockData to/from')
# 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')
return parser
def _add_zero_args(parser):
"""Text generate arguments."""
group = parser.add_argument_group('Text generation', 'configurations')
group.add_argument("--zero-stage", type=int, default=1.0)
group.add_argument('--zero-reduce-scatter', action='store_true',
help='Use reduce scatter if specified')
group.add_argument('--zero-contigious-gradients', action='store_true',
help='Use contigious memory optimizaiton if specified')
group.add_argument("--zero-reduce-bucket-size", type=int, default=0.0)
group.add_argument("--zero-allgather-bucket-size", type=int, default=0.0)
return parser
def _add_activation_checkpoint_args(parser):
group = parser.add_argument_group('Activation Checkpointing',
'Checkpointing Configurations')
group.add_argument('--deepspeed-activation-checkpointing', action='store_true',
help='uses activation checkpointing from deepspeed')
group.add_argument('--partition-activations', action='store_true',
help='partition Activations across GPUs before checkpointing.')
group.add_argument('--contigious-checkpointing', action='store_true',
help='Contigious memory checkpointing for activatoins.')
group.add_argument('--checkpoint-in-cpu', action='store_true',
help='Move the activation checkpoints to CPU.')
group.add_argument('--synchronize-each-layer', action='store_true',
help='does a synchronize at the beginning and end of each checkpointed layer.')
group.add_argument('--profile-backward', action='store_true',
help='Enables backward pass profiling for checkpointed layers.')
return parser