training.py 38.1 KB
Newer Older
1
# coding=utf-8
Mohammad's avatar
Mohammad committed
2
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
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
"""Pretrain utilities."""
17
18
19

from datetime import datetime
import math
Mohammad's avatar
Mohammad committed
20
import sys
21
22
23
24
import time
# The earliest we can measure the start time.
_TRAIN_START_TIME = time.time()

25
26
27
import torch
from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP

Neel Kant's avatar
Neel Kant committed
28
from megatron import get_args
29
from megatron import get_signal_handler
Mohammad's avatar
Mohammad committed
30
31
from megatron import get_timers
from megatron import get_tensorboard_writer
32
from megatron import get_current_global_batch_size
mohammad's avatar
mohammad committed
33
from megatron import get_num_microbatches
mohammad's avatar
mohammad committed
34
from megatron import is_last_rank
mohammad's avatar
mohammad committed
35
from megatron import update_num_microbatches
36
from megatron import mpu
Neel Kant's avatar
Neel Kant committed
37
from megatron import print_rank_0
38
from megatron import print_rank_last
Mohammad's avatar
Mohammad committed
39
40
from megatron.checkpointing import load_checkpoint
from megatron.checkpointing import save_checkpoint
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
41
from megatron.model import Float16Module
42
from megatron.model import ModelType
mohammad's avatar
mohammad committed
43
from megatron.optimizer import get_megatron_optimizer
Mohammad's avatar
Mohammad committed
44
from megatron.initialize import initialize_megatron
45
from megatron.initialize import write_args_to_tensorboard
46
47
48
from megatron.learning_rates import AnnealingLR
from megatron.model import DistributedDataParallel as LocalDDP
from megatron.utils import check_adlr_autoresume_termination
49
from megatron.utils import unwrap_model
Vijay Korthikanti's avatar
Vijay Korthikanti committed
50
from megatron.data.data_samplers import build_pretraining_data_loader
mohammad's avatar
mohammad committed
51
from megatron.utils import calc_params_l2_norm
52
from megatron.schedules import get_forward_backward_func
53
from megatron.utils import report_memory
54
55


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
56

57
58
59
60
61
62
63
def print_datetime(string):
    """Note that this call will sync across all ranks."""
    torch.distributed.barrier()
    time_str = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
    print_rank_0('[' + string + '] datetime: {} '.format(time_str))


64
def pretrain(train_valid_test_dataset_provider,
65
             model_provider,
66
             model_type,
67
68
             forward_step_func,
             extra_args_provider=None,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
69
             args_defaults={}):
70
71
72
    """Main training program.

    This function will run the followings in the order provided:
Mohammad's avatar
Mohammad committed
73
74
        1) initialize Megatron.
        2) setup model, optimizer and lr schedule using the model_provider.
75
        3) call train_val_test_data_provider to get train/val/test datasets.
Mohammad's avatar
Mohammad committed
76
        4) train the modle using the forward_step_func.
77
78

    Arguments:
79
80
81
        train_valid_test_dataset_provider: a function that takes the size of
            train/valid/test dataset and returns `train, valid, test` datasets.
        model_provider: a function that returns a vanilla version of the
Mohammad's avatar
Mohammad committed
82
            model. By vanilla we mean a simple model on cpu with no fp16 or ddp.
83
        model_type: an enum that specifies the type of model being trained.
Mohammad's avatar
Mohammad committed
84
85
86
87
88
89
90
91
92
        forward_step_func: a function that takes a `data iterator` and `model`,
            and returns a `loss` scalar with a dictionary with key:values being
            the info we would like to monitor during training, for example
            `lm-loss: value`. We also require that this function add
            `batch generator` to the timers class.
        extra_args_provider: a function that takes a parser and adds arguments
            to it. It is used for programs to add their own arguments.
        args_defaults: a dictionary from argument-name to argument-value. It
            to set already parse arguments.
93
94
    """

95
    # Initalize and get arguments, timers, and Tensorboard writer.
96
97
    initialize_megatron(extra_args_provider=extra_args_provider,
                        args_defaults=args_defaults)
98

99
100
101
102
    # Adjust the startup time so it reflects the largest value.
    # This will be closer to what scheduler will see (outside of
    # image ... launches.
    global _TRAIN_START_TIME
103
    start_time_tensor = torch.cuda.DoubleTensor([_TRAIN_START_TIME])
104
105
106
    torch.distributed.all_reduce(start_time_tensor,
                                 op=torch.distributed.ReduceOp.MIN)
    _TRAIN_START_TIME = start_time_tensor.item()
mshoeybi's avatar
mshoeybi committed
107
    print_rank_0('time to initialize megatron (seconds): {:.3f}'.format(
108
109
110
        time.time() - _TRAIN_START_TIME))
    print_datetime('after megatron is initialized')

111
    args = get_args()
Mohammad's avatar
Mohammad committed
112
    timers = get_timers()
113
114

    # Model, optimizer, and learning rate.
115
    timers('model-and-optimizer-setup').start()
116
117
    model, optimizer, lr_scheduler = setup_model_and_optimizer(model_provider,
                                                               model_type)
118
    timers('model-and-optimizer-setup').stop()
119
120
    print_datetime('after model, optimizer, and learning rate '
                   'scheduler are built')
121
122

    # Data stuff.
123
124
    timers('train/valid/test-data-iterators-setup').start()
    if args.virtual_pipeline_model_parallel_size is not None:
125
        all_data_iterators = [
126
127
128
            build_train_valid_test_data_iterators(train_valid_test_dataset_provider)
            for _ in range(len(model))
        ]
129
130
131
        train_data_iterator = [data_iterators[0] for data_iterators in all_data_iterators]
        valid_data_iterator = [data_iterators[1] for data_iterators in all_data_iterators]
        test_data_iterator = [data_iterators[2] for data_iterators in all_data_iterators]
132
133
134
135
136
    else:
        train_data_iterator, valid_data_iterator, test_data_iterator \
            = build_train_valid_test_data_iterators(
                train_valid_test_dataset_provider)
    timers('train/valid/test-data-iterators-setup').stop()
mshoeybi's avatar
mshoeybi committed
137
    print_datetime('after dataloaders are built')
Mohammad's avatar
Mohammad committed
138
139

    # Print setup timing.
140
141
    print_rank_0('done with setup ...')
    timers.log(['model-and-optimizer-setup', 'train/valid/test-data-iterators-setup'])
Mohammad's avatar
Mohammad committed
142
    print_rank_0('training ...')
143
144

    iteration = 0
145
    if args.do_train and args.train_iters > 0:
mohammad's avatar
mohammad committed
146
147
148
        iteration = train(forward_step_func,
                          model, optimizer, lr_scheduler,
                          train_data_iterator, valid_data_iterator)
149
    print_datetime('after training is done')
Mohammad's avatar
Mohammad committed
150

151
152
153
    if args.do_valid:
        prefix = 'the end of training for val data'
        evaluate_and_print_results(prefix, forward_step_func,
154
                                   valid_data_iterator, model,
Mohammad's avatar
Mohammad committed
155
                                   iteration, False)
156
157

    if args.save and iteration != 0:
158
        save_checkpoint(iteration, model, optimizer, lr_scheduler)
159
160
161
162
163
164

    if args.do_test:
        # Run on test data.
        prefix = 'the end of training for test data'
        evaluate_and_print_results(prefix, forward_step_func,
                                   test_data_iterator, model,
Mohammad's avatar
Mohammad committed
165
                                   0, True)
166

167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
def update_train_iters(args):

    # For iteration-based training, we don't need to do anything
    if args.train_iters:
        return

    # Constant batch size with sample-based training.
    if args.rampup_batch_size is None:
        args.train_iters = args.train_samples // args.global_batch_size

    else:
        # Sample based training with rampup batch size.
        iterations = 0
        consumed_samples = 0
        # Rampup phase.
        while consumed_samples <= int(args.rampup_batch_size[2]):
183
184
            update_num_microbatches(consumed_samples, consistency_check=False)
            consumed_samples += get_current_global_batch_size()
185
186
            iterations += 1
        # Reset
187
        update_num_microbatches(0, consistency_check=False)
188
189
190
191
192
193
194
195
        # Constant phase
        # Note that we throw away any partial last batch.
        iterations += (args.train_samples - consumed_samples) // \
                      args.global_batch_size
        args.train_iters = iterations

    print_rank_0('setting training iterations to {}'.format(args.train_iters))

196

197
def get_model(model_provider_func, model_type=ModelType.encoder_or_decoder, wrap_with_ddp=True):
198
    """Build the model."""
Mohammad's avatar
Mohammad committed
199
    args = get_args()
200
    args.model_type = model_type
201

202
    # Build model.
203
204
    if mpu.get_pipeline_model_parallel_world_size() > 1 and \
       args.virtual_pipeline_model_parallel_size is not None:
205
206
        assert model_type != ModelType.encoder_and_decoder, \
            "Interleaved schedule not supported for model with both encoder and decoder"
207
208
209
        model = []
        for i in range(args.virtual_pipeline_model_parallel_size):
            mpu.set_virtual_pipeline_model_parallel_rank(i)
210
211
212
            # Set pre_process and post_process only after virtual rank is set.
            pre_process = mpu.is_pipeline_first_stage()
            post_process = mpu.is_pipeline_last_stage()
213
            this_model = model_provider_func(
214
215
216
                pre_process=pre_process,
                post_process=post_process
            )
217
            this_model.model_type = model_type
218
            model.append(this_model)
219
    else:
220
221
        pre_process = mpu.is_pipeline_first_stage()
        post_process = mpu.is_pipeline_last_stage()
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
        add_encoder = True
        add_decoder = True
        if model_type == ModelType.encoder_and_decoder:
            if mpu.get_pipeline_model_parallel_world_size() > 1:
                assert args.pipeline_model_parallel_split_rank is not None, \
                    "Split rank needs to be specified for model with both encoder and decoder"
                rank = mpu.get_pipeline_model_parallel_rank()
                split_rank = args.pipeline_model_parallel_split_rank
                world_size = mpu.get_pipeline_model_parallel_world_size()
                pre_process = rank == 0 or rank == split_rank
                post_process = (rank == (split_rank - 1)) or (
                        rank == (world_size - 1))
                add_encoder = mpu.is_pipeline_stage_before_split()
                add_decoder = mpu.is_pipeline_stage_after_split()
            model = model_provider_func(
                pre_process=pre_process,
                post_process=post_process,
                add_encoder=add_encoder,
                add_decoder=add_decoder)
        else:
            model = model_provider_func(
                pre_process=pre_process,
                post_process=post_process
            )
        model.model_type = model_type
247

248
249
    if not isinstance(model, list):
        model = [model]
250

251
    # Set tensor model parallel attributes if not set.
mohammad's avatar
mohammad committed
252
253
254
    # Only parameters that are already tensor model parallel have these
    # attributes set for them. We should make sure the default attributes
    # are set for all params so the optimizer can use them.
255
256
257
    for model_module in model:
        for param in model_module.parameters():
            mpu.set_defaults_if_not_set_tensor_model_parallel_attributes(param)
258

259
260
    # Print number of parameters.
    if mpu.get_data_parallel_rank() == 0:
261
        print(' > number of parameters on (tensor, pipeline) '
262
              'model parallel rank ({}, {}): {}'.format(
263
264
            mpu.get_tensor_model_parallel_rank(),
            mpu.get_pipeline_model_parallel_rank(),
265
266
            sum([sum([p.nelement() for p in model_module.parameters()])
                 for model_module in model])), flush=True)
267
268

    # GPU allocation.
269
270
    for model_module in model:
        model_module.cuda(torch.cuda.current_device())
271
272

    # Fp16 conversion.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
273
274
    if args.fp16 or args.bf16:
        model = [Float16Module(model_module, args) for model_module in model]
275

276
277
278
279
280
281
    if wrap_with_ddp:
        if args.DDP_impl == 'torch':
            i = torch.cuda.current_device()
            model = [torchDDP(model_module, device_ids=[i], output_device=i,
                              process_group=mpu.get_data_parallel_group())
                     for model_module in model]
282

283
284
285
286
287
288
289
290
291
        elif args.DDP_impl == 'local':
            model = [LocalDDP(model_module,
                              args.accumulate_allreduce_grads_in_fp32,
                              args.use_contiguous_buffers_in_local_ddp)
                     for model_module in model]

        else:
            raise NotImplementedError('Unknown DDP implementation specified: '
                                      '{}. Exiting.'.format(args.DDP_impl))
292

293
    return model
294
295


Mohammad's avatar
Mohammad committed
296
def get_learning_rate_scheduler(optimizer):
297
    """Build the learning rate scheduler."""
Mohammad's avatar
Mohammad committed
298
    args = get_args()
299

300
301
302
303
304
    # Iteration-based training.
    if args.train_iters:
        if args.lr_decay_iters is None:
            args.lr_decay_iters = args.train_iters
        decay_steps = args.lr_decay_iters * args.global_batch_size
305
306
        if args.lr_warmup_fraction is not None:
            warmup_steps = args.lr_warmup_fraction * decay_steps
307
308
        else:
            warmup_steps = args.lr_warmup_iters * args.global_batch_size
309
310
311
312
313
    # Sample-based training.
    elif args.train_samples:
        # We need to set training iters for later use. Technically
        # we need to adjust the training samples too (due to last
        # batch being incomplete) but we leave it as is for now.
314
        update_train_iters(args)
315
316
317
        if args.lr_decay_samples is None:
            args.lr_decay_samples = args.train_samples
        decay_steps = args.lr_decay_samples
318
319
        if args.lr_warmup_fraction is not None:
            warmup_steps = args.lr_warmup_fraction * decay_steps
320
321
        else:
            warmup_steps = args.lr_warmup_samples
322
    else:
323
324
325
        raise Exception(
            'either train-iters or train-samples should be provided.')

326
327
    lr_scheduler = AnnealingLR(
        optimizer,
328
        max_lr=args.lr,
329
        min_lr=args.min_lr,
330
331
        warmup_steps=warmup_steps,
        decay_steps=decay_steps,
332
        decay_style=args.lr_decay_style,
333
334
335
336
337
338
        use_checkpoint_lr_scheduler=args.use_checkpoint_lr_scheduler,
        override_lr_scheduler=args.override_lr_scheduler)

    return lr_scheduler


339
def setup_model_and_optimizer(model_provider_func, model_type):
340
    """Setup model and optimizer."""
Mohammad's avatar
Mohammad committed
341
    args = get_args()
342

343
    model = get_model(model_provider_func, model_type)
344

345
    unwrapped_model = unwrap_model(model,
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
346
                                   (torchDDP, LocalDDP, Float16Module))
347
348
    optimizer = get_megatron_optimizer(unwrapped_model)

Mohammad's avatar
Mohammad committed
349
    lr_scheduler = get_learning_rate_scheduler(optimizer)
350
351

    if args.load is not None:
352
353
354
355
        timers = get_timers()
        # Extra barrier is added to make sure all ranks report the
        # max time.
        torch.distributed.barrier()
356
        timers('load-checkpoint').start()
357
        args.iteration = load_checkpoint(model, optimizer, lr_scheduler)
358
        torch.distributed.barrier()
359
360
        timers('load-checkpoint').stop()
        timers.log(['load-checkpoint'])
361
362
363
    else:
        args.iteration = 0

mohammad's avatar
mohammad committed
364
    # We only support local DDP with multiple micro-batches.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
365
    if len(model) > 1 or mpu.get_pipeline_model_parallel_world_size() > 1:
mohammad's avatar
mohammad committed
366
367
        assert args.DDP_impl == 'local'

Neel Kant's avatar
Neel Kant committed
368
    # get model without FP16 and/or TorchDDP wrappers
Mostofa Patwary's avatar
Mostofa Patwary committed
369
370
    if args.iteration == 0 and len(unwrapped_model) == 1 \
        and hasattr(unwrapped_model[0], 'init_state_dict_from_bert'):
Mostofa Patwary's avatar
Mostofa Patwary committed
371
        print_rank_0("Initializing ICT from pretrained BERT model")
Mostofa Patwary's avatar
Mostofa Patwary committed
372
        unwrapped_model[0].init_state_dict_from_bert()
Mostofa Patwary's avatar
Mostofa Patwary committed
373
374
        if args.fp16:
            optimizer.reload_model_params()
Neel Kant's avatar
Neel Kant committed
375

376
377
378
    return model, optimizer, lr_scheduler


379
380
381
382
383
384
385
def train_step(forward_step_func, data_iterator,
               model, optimizer, lr_scheduler):
    """Single training step."""
    args = get_args()
    timers = get_timers()

    # Set grad to zero.
386
    if args.DDP_impl == 'local' and args.use_contiguous_buffers_in_local_ddp:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
387
388
        for partition in model:
            partition.zero_grad_buffer()
389
    optimizer.zero_grad()
390

391
    forward_backward_func = get_forward_backward_func()
392
393
394
    losses_reduced = forward_backward_func(
        forward_step_func, data_iterator, model,
        optimizer, timers, forward_only=False)
395

396
    # Empty unused memory
Lawrence McAfee's avatar
Lawrence McAfee committed
397
    if args.empty_unused_memory_level >= 1:
398
399
        torch.cuda.empty_cache()

400
401
    # All-reduce if needed.
    if args.DDP_impl == 'local':
402
        timers('backward-params-all-reduce').start()
403
        for model_module in model:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
404
            model_module.allreduce_gradients()
405
        timers('backward-params-all-reduce').stop()
406

407
408
409
410
    # All-reduce word_embeddings' grad across first and last stages to ensure
    # that word_embeddings parameters stay in sync.
    # This should only run for models that support pipelined model parallelism
    # (BERT and GPT-2).
411
    timers('backward-embedding-all-reduce').start()
412
    if mpu.is_rank_in_embedding_group(ignore_virtual=True) and \
413
            mpu.get_pipeline_model_parallel_world_size() > 1:
414
415
416
417
        if mpu.is_pipeline_first_stage(ignore_virtual=True):
            unwrapped_model = model[0]
        elif mpu.is_pipeline_last_stage(ignore_virtual=True):
            unwrapped_model = model[-1]
418
419
        else:  # We do not support the interleaved schedule for T5 yet.
            unwrapped_model = model[0]
420
        unwrapped_model = unwrap_model(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
421
            unwrapped_model, (torchDDP, LocalDDP, Float16Module))
422

423
424
        if unwrapped_model.share_word_embeddings:
            word_embeddings_weight = unwrapped_model.word_embeddings_weight()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
425
426
427
428
429
            if args.DDP_impl == 'local':
                grad = word_embeddings_weight.main_grad
            else:
                grad = word_embeddings_weight.grad
            torch.distributed.all_reduce(grad, group=mpu.get_embedding_group())
Vijay Korthikanti's avatar
Vijay Korthikanti committed
430
431
432
433
434
435
436
437
438

    if mpu.is_rank_in_position_embedding_group() and \
            mpu.get_pipeline_model_parallel_world_size() > 1 and \
            args.pipeline_model_parallel_split_rank is not None:
        unwrapped_model = model[0]
        unwrapped_model = unwrap_model(
            unwrapped_model, (torchDDP, LocalDDP, Float16Module))
        grad = unwrapped_model.language_model.embedding.position_embeddings.weight.main_grad
        torch.distributed.all_reduce(grad, group=mpu.get_position_embedding_group())
439
    timers('backward-embedding-all-reduce').stop()
440

441
442
    # Update parameters.
    timers('optimizer').start()
443
    update_successful, grad_norm, num_zeros_in_grad = optimizer.step()
444
445
446
    timers('optimizer').stop()

    # Update learning rate.
447
    if update_successful:
448
449
450
451
        increment = get_num_microbatches() * \
                    args.micro_batch_size * \
                    args.data_parallel_size
        lr_scheduler.step(increment=increment)
mohammad's avatar
mohammad committed
452
        skipped_iter = 0
453
454
455
    else:
        skipped_iter = 1

456
    # Empty unused memory
Lawrence McAfee's avatar
Lawrence McAfee committed
457
    if args.empty_unused_memory_level >= 2:
458
459
        torch.cuda.empty_cache()

460
    if mpu.is_pipeline_last_stage(ignore_virtual=True):
461
462
463
464
        # Average loss across microbatches.
        loss_reduced = {}
        for key in losses_reduced[0]:
            losses_reduced_for_key = [x[key] for x in losses_reduced]
465
            loss_reduced[key] = sum(losses_reduced_for_key) / len(losses_reduced_for_key)
466
467
        return loss_reduced, skipped_iter, grad_norm, num_zeros_in_grad
    return {}, skipped_iter, grad_norm, num_zeros_in_grad
468
469


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
470
def training_log(loss_dict, total_loss_dict, learning_rate, iteration,
mohammad's avatar
mohammad committed
471
                 loss_scale, report_memory_flag, skipped_iter,
472
                 grad_norm, params_norm, num_zeros_in_grad):
Mohammad's avatar
Mohammad committed
473
474
475
476
    """Log training information such as losses, timing, ...."""
    args = get_args()
    timers = get_timers()
    writer = get_tensorboard_writer()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
477

mohammad's avatar
mohammad committed
478
479
    # Advanced, skipped, and Nan iterations.
    advanced_iters_key = 'advanced iterations'
mohammad's avatar
mohammad committed
480
    skipped_iters_key = 'skipped iterations'
mohammad's avatar
mohammad committed
481
482
483
484
485
486
487
488
489
    nan_iters_key = 'nan iterations'
    # Advanced iterations.
    if not skipped_iter:
        total_loss_dict[advanced_iters_key] = total_loss_dict.get(
            advanced_iters_key, 0) + 1
    else:
        if advanced_iters_key not in total_loss_dict:
            total_loss_dict[advanced_iters_key] = 0
    # Skipped iterations.
mohammad's avatar
mohammad committed
490
491
    total_loss_dict[skipped_iters_key] = total_loss_dict.get(
        skipped_iters_key, 0) + skipped_iter
mohammad's avatar
mohammad committed
492
    # Update losses and set nan iterations
mohammad's avatar
mohammad committed
493
    got_nan = False
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
494
    for key in loss_dict:
mohammad's avatar
mohammad committed
495
        if not skipped_iter:
496
497
            total_loss_dict[key] = total_loss_dict.get(
                key, torch.cuda.FloatTensor([0.0])) + loss_dict[key]
mohammad's avatar
mohammad committed
498
499
500
501
502
        else:
            value = loss_dict[key].float().sum().item()
            is_nan = value == float('inf') or \
                     value == -float('inf') or \
                     value != value
mohammad's avatar
mohammad committed
503
            got_nan = got_nan or is_nan
mohammad's avatar
mohammad committed
504
505
    total_loss_dict[nan_iters_key] = total_loss_dict.get(
        nan_iters_key, 0) + int(got_nan)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
506
507
508

    # Logging.
    timers_to_log = []
Neel Kant's avatar
Neel Kant committed
509

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
510
511
512
    def add_to_logging(name):
        if name in timers.timers:
            timers_to_log.append(name)
513
514
515
    add_to_logging('forward-compute')
    add_to_logging('forward-recv')
    add_to_logging('forward-send')
516
    add_to_logging('forward-backward-send-forward-backward-recv')
517
518
519
    add_to_logging('backward-compute')
    add_to_logging('backward-recv')
    add_to_logging('backward-send')
Deepak Narayanan's avatar
Deepak Narayanan committed
520
    add_to_logging('backward-send-forward-recv')
521
    add_to_logging('backward-send-backward-recv')
522
    add_to_logging('backward-params-all-reduce')
523
    add_to_logging('backward-embedding-all-reduce')
524
    add_to_logging('optimizer-copy-to-main-grad')
mohammad's avatar
mohammad committed
525
    add_to_logging('optimizer-unscale-and-check-inf')
526
527
    add_to_logging('optimizer-clip-main-grad')
    add_to_logging('optimizer-copy-main-to-model-params')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
528
    add_to_logging('optimizer')
mohammad's avatar
mohammad committed
529
    add_to_logging('batch-generator')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
530

mohammad's avatar
mohammad committed
531
    # Calculate batch size.
mshoeybi's avatar
mshoeybi committed
532
533
534
    batch_size = args.micro_batch_size * args.data_parallel_size * \
        get_num_microbatches()

mohammad's avatar
mohammad committed
535
536
537
    total_iterations = total_loss_dict[advanced_iters_key] + \
                       total_loss_dict[skipped_iters_key]

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
538
    # Tensorboard values.
539
540
541
542
543
544
545
546
547
548
    if writer and (iteration % args.tensorboard_log_interval == 0 ) and \
       is_last_rank():
        if args.log_learning_rate_to_tensorboard:
            writer.add_scalar('learning-rate', learning_rate, iteration)
            writer.add_scalar('learning-rate vs samples', learning_rate,
                              args.consumed_train_samples)
        if args.log_batch_size_to_tensorboard:
            writer.add_scalar('batch-size', batch_size, iteration)
            writer.add_scalar('batch-size vs samples', batch_size,
                              args.consumed_train_samples)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
549
        for key in loss_dict:
mohammad's avatar
mohammad committed
550
551
            writer.add_scalar(key , loss_dict[key], iteration)
            writer.add_scalar(key + ' vs samples', loss_dict[key],
552
                              args.consumed_train_samples)
553
554
555
556
        if args.log_loss_scale_to_tensorboard:
            writer.add_scalar('loss-scale', loss_scale, iteration)
            writer.add_scalar('loss-scale vs samples', loss_scale,
                              args.consumed_train_samples)
557
558
559
560
        if args.log_world_size_to_tensorboard:
            writer.add_scalar('world-size', args.world_size, iteration)
            writer.add_scalar('world-size vs samples', args.world_size,
                              args.consumed_train_samples)
561
562
563
564
        if grad_norm is not None:
            writer.add_scalar('grad-norm', grad_norm, iteration)
            writer.add_scalar('grad-norm vs samples', grad_norm,
                              args.consumed_train_samples)
565
566
567
        if num_zeros_in_grad is not None:
            writer.add_scalar('num-zeros', num_zeros_in_grad, iteration)
            writer.add_scalar('num-zeros vs samples', num_zeros_in_grad,
Rewon Child's avatar
Rewon Child committed
568
                              args.consumed_train_samples)
mohammad's avatar
mohammad committed
569
570
571
572
        if params_norm is not None:
            writer.add_scalar('params-norm', params_norm, iteration)
            writer.add_scalar('params-norm vs samples', params_norm,
                              args.consumed_train_samples)
573
574
575
        if args.log_timers_to_tensorboard:
            timers.write(timers_to_log, writer, iteration,
                         normalizer=total_iterations)
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
        if args.log_memory_to_tensorboard:
            mem_stats = torch.cuda.memory_stats()
            writer.add_scalar(
                "mem-reserved-bytes",
                mem_stats["reserved_bytes.all.current"],
                iteration,
            )
            writer.add_scalar(
                "mem-allocated-bytes",
                mem_stats["allocated_bytes.all.current"],
                iteration,
            )
            writer.add_scalar(
                "mem-allocated-count",
                mem_stats["allocation.all.current"],
                iteration,
            )
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
593
594

    if iteration % args.log_interval == 0:
595
        elapsed_time = timers('interval-time').elapsed()
mohammad's avatar
mohammad committed
596
        elapsed_time_per_iteration = elapsed_time / total_iterations
mshoeybi's avatar
mshoeybi committed
597
        if writer:
598
599
600
            if args.log_timers_to_tensorboard:
                writer.add_scalar('iteration-time',
                                  elapsed_time_per_iteration, iteration)
601
602
        log_string = ' iteration {:8d}/{:8d} |'.format(
            iteration, args.train_iters)
mshoeybi's avatar
mshoeybi committed
603
        log_string += ' consumed samples: {:12d} |'.format(
604
            args.consumed_train_samples)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
605
        log_string += ' elapsed time per iteration (ms): {:.1f} |'.format(
mohammad's avatar
mohammad committed
606
            elapsed_time_per_iteration * 1000.0)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
607
        log_string += ' learning rate: {:.3E} |'.format(learning_rate)
mohammad's avatar
mohammad committed
608
        log_string += ' global batch size: {:5d} |'.format(batch_size)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
609
        for key in total_loss_dict:
mohammad's avatar
mohammad committed
610
611
612
613
            if key not in [advanced_iters_key, skipped_iters_key,
                           nan_iters_key]:
                avg = total_loss_dict[key].item() / \
                      float(max(1, total_loss_dict[advanced_iters_key]))
614
615
616
                if avg > 0.0:
                    log_string += ' {}: {:.6E} |'.format(key, avg)
                total_loss_dict[key] = torch.cuda.FloatTensor([0.0])
617
        log_string += ' loss scale: {:.1f} |'.format(loss_scale)
618
619
        if grad_norm is not None:
            log_string += ' grad norm: {:.3f} |'.format(grad_norm)
620
621
        if num_zeros_in_grad is not None:
            log_string += ' num zeros: {:.1f} |'.format(num_zeros_in_grad)
mohammad's avatar
mohammad committed
622
623
        if params_norm is not None:
            log_string += ' params norm: {:.3f} |'.format(params_norm)
mohammad's avatar
mohammad committed
624
625
        log_string += ' number of skipped iterations: {:3d} |'.format(
            total_loss_dict[skipped_iters_key])
mohammad's avatar
mohammad committed
626
        log_string += ' number of nan iterations: {:3d} |'.format(
mohammad's avatar
mohammad committed
627
628
            total_loss_dict[nan_iters_key])
        total_loss_dict[advanced_iters_key] = 0
mohammad's avatar
mohammad committed
629
        total_loss_dict[skipped_iters_key] = 0
mohammad's avatar
mohammad committed
630
        total_loss_dict[nan_iters_key] = 0
631
        print_rank_last(log_string)
632
633
634
        if report_memory_flag and learning_rate > 0.:
            # Report memory after optimizer state has been initialized.
            report_memory('(after {} iterations)'.format(iteration))
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
635
636
637
638
639
640
            report_memory_flag = False
        timers.log(timers_to_log, normalizer=args.log_interval)

    return report_memory_flag


641
642
643
644
645
def save_checkpoint_and_time(iteration, model, optimizer, lr_scheduler):
    timers = get_timers()
    # Extra barrier is added to make sure
    # all ranks report the max time.
    torch.distributed.barrier()
646
    timers('save-checkpoint').start()
647
648
    save_checkpoint(iteration, model, optimizer, lr_scheduler)
    torch.distributed.barrier()
649
650
    timers('save-checkpoint').stop()
    timers.log(['save-checkpoint'])
651
652


653
def train(forward_step_func, model, optimizer, lr_scheduler,
654
          train_data_iterator, valid_data_iterator):
655
    """Train the model function."""
Mohammad's avatar
Mohammad committed
656
657
    args = get_args()
    timers = get_timers()
658

659
660
661
    # Write args to tensorboard
    write_args_to_tensorboard()

662
    # Turn on training mode which enables dropout.
663
664
    for model_module in model:
        model_module.train()
665
666
667
668
669
670
671

    # Tracking loss.
    total_loss_dict = {}

    # Iterations.
    iteration = args.iteration

672
    timers('interval-time').start()
673
    print_datetime('before the start of training step')
674
675
    report_memory_flag = True
    while iteration < args.train_iters:
mohammad's avatar
mohammad committed
676
        update_num_microbatches(args.consumed_train_samples)
677
678
679
680
681
682
        loss_dict, skipped_iter, grad_norm, num_zeros_in_grad = \
            train_step(forward_step_func,
                       train_data_iterator,
                       model,
                       optimizer,
                       lr_scheduler)
683
        iteration += 1
684
        args.consumed_train_samples += mpu.get_data_parallel_world_size() * \
685
                                       args.micro_batch_size * \
mohammad's avatar
mohammad committed
686
                                       get_num_microbatches()
687
688

        # Logging.
689
        loss_scale = optimizer.get_loss_scale().item()
690
691
692
        params_norm = None
        if args.log_params_norm:
            params_norm = calc_params_l2_norm(model)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
693
694
        report_memory_flag = training_log(loss_dict, total_loss_dict,
                                          optimizer.param_groups[0]['lr'],
Mohammad's avatar
Mohammad committed
695
                                          iteration, loss_scale,
696
                                          report_memory_flag, skipped_iter,
697
                                          grad_norm, params_norm, num_zeros_in_grad)
698
699

        # Autoresume
700
701
        if args.adlr_autoresume and \
           (iteration % args.adlr_autoresume_interval == 0):
702
            check_adlr_autoresume_termination(iteration, model, optimizer,
703
                                              lr_scheduler)
704
705
706
707
708
709

        # Evaluation
        if args.eval_interval and iteration % args.eval_interval == 0 and \
           args.do_valid:
            prefix = 'iteration {}'.format(iteration)
            evaluate_and_print_results(prefix, forward_step_func,
710
                                       valid_data_iterator, model,
Mohammad's avatar
Mohammad committed
711
                                       iteration, False)
712

713
714
        # Checkpointing
        saved_checkpoint = False
715
716
717
718
719
720
721
722
        if args.exit_signal_handler:
            signal_handler = get_signal_handler()
            if any(signal_handler.signals_received()):
                save_checkpoint_and_time(iteration, model, optimizer,
                                         lr_scheduler)
                print_datetime('exiting program after receiving SIGTERM.')
                sys.exit()

723
724
725
726
727
728
        if args.save and args.save_interval and \
           iteration % args.save_interval == 0:
            save_checkpoint_and_time(iteration, model, optimizer,
                                     lr_scheduler)
            saved_checkpoint = True

729
730
731
732
733
734
735
736
737
738
739
740
        # Exiting based on duration
        if args.exit_duration_in_mins:
            train_time = (time.time() - _TRAIN_START_TIME) / 60.0
            done_cuda = torch.cuda.IntTensor(
                [train_time > args.exit_duration_in_mins])
            torch.distributed.all_reduce(
                done_cuda, op=torch.distributed.ReduceOp.MAX)
            done = done_cuda.item()
            if done:
                if not saved_checkpoint:
                    save_checkpoint_and_time(iteration, model, optimizer,
                                             lr_scheduler)
741
                print_datetime('exiting program after {} minutes'.format(train_time))
742
743
                sys.exit()

744
        # Exiting based on iterations
745
        if args.exit_interval and iteration % args.exit_interval == 0:
746
747
748
            if not saved_checkpoint:
                save_checkpoint_and_time(iteration, model, optimizer,
                                         lr_scheduler)
749
            torch.distributed.barrier()
750
            print_datetime('exiting program at iteration {}'.format(iteration))
Mohammad's avatar
Mohammad committed
751
            sys.exit()
752

753

mohammad's avatar
mohammad committed
754
    return iteration
755
756


Mohammad's avatar
Mohammad committed
757
def evaluate(forward_step_func, data_iterator, model, verbose=False):
758
    """Evaluation."""
Mohammad's avatar
Mohammad committed
759
    args = get_args()
760
761

    # Turn on evaluation mode which disables dropout.
762
763
    for model_module in model:
        model_module.eval()
764
765
766
767
768
769
770
771
772
773

    total_loss_dict = {}

    with torch.no_grad():
        iteration = 0
        while iteration < args.eval_iters:
            iteration += 1
            if verbose and iteration % args.log_interval == 0:
                print_rank_0('Evaluating iter {}/{}'.format(iteration,
                                                            args.eval_iters))
774

775
            forward_backward_func = get_forward_backward_func()
776
777
778
779
            loss_dicts = forward_backward_func(
                forward_step_func, data_iterator, model, optimizer=None,
                timers=None, forward_only=True)

780
            # Empty unused memory
Lawrence McAfee's avatar
Lawrence McAfee committed
781
            if args.empty_unused_memory_level >= 1:
782
783
                torch.cuda.empty_cache()

784
785
786
            if mpu.is_pipeline_last_stage(ignore_virtual=True):
                # Reduce across processes.
                for loss_dict in loss_dicts:
787
                    for key in loss_dict:
788
789
                        total_loss_dict[key] = total_loss_dict.get(
                            key, torch.cuda.FloatTensor([0.0])) + loss_dict[key]
790

791
            args.consumed_valid_samples += mpu.get_data_parallel_world_size() \
792
                                           * args.micro_batch_size \
mohammad's avatar
mohammad committed
793
                                           * get_num_microbatches()
794
    # Move model back to the train mode.
795
796
    for model_module in model:
        model_module.train()
797
798

    for key in total_loss_dict:
mohammad's avatar
mohammad committed
799
        total_loss_dict[key] /= args.eval_iters * get_num_microbatches()
800
801
802
803
804

    return total_loss_dict

def evaluate_and_print_results(prefix, forward_step_func,
                               data_iterator, model,
Mohammad's avatar
Mohammad committed
805
                               iteration, verbose=False):
806
    """Helper function to evaluate and dump results on screen."""
807
    args = get_args()
Mohammad's avatar
Mohammad committed
808
809
810
    writer = get_tensorboard_writer()

    total_loss_dict = evaluate(forward_step_func, data_iterator, model, verbose)
811
812
813
814
815
    string = ' validation loss at {} | '.format(prefix)
    for key in total_loss_dict:
        string += '{} value: {:.6E} | '.format(key, total_loss_dict[key].item())
        ppl = math.exp(min(20, total_loss_dict[key].item()))
        string += '{} PPL: {:.6E} | '.format(key, ppl)
mshoeybi's avatar
mshoeybi committed
816
        if writer:
mohammad's avatar
mohammad committed
817
            writer.add_scalar('{} validation'.format(key),
818
819
                              total_loss_dict[key].item(),
                              iteration)
mohammad's avatar
mohammad committed
820
            writer.add_scalar('{} validation vs samples'.format(key),
821
822
                              total_loss_dict[key].item(),
                              args.consumed_train_samples)
823
            if args.log_validation_ppl_to_tensorboard:
mohammad's avatar
mohammad committed
824
                writer.add_scalar('{} validation ppl'.format(key), ppl,
825
                                  iteration)
mohammad's avatar
mohammad committed
826
                writer.add_scalar('{} validation ppl vs samples'.format(key),
827
                                  ppl, args.consumed_train_samples)
828
829

    length = len(string) + 1
830
831
832
    print_rank_last('-' * length)
    print_rank_last(string)
    print_rank_last('-' * length)
833
834


Vijay Korthikanti's avatar
Vijay Korthikanti committed
835
def cyclic_iter(iter):
836
    while True:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
837
        for x in iter:
838
839
            yield x

840
841
842
def build_train_valid_test_data_iterators(
        build_train_valid_test_datasets_provider):
    """XXX"""
Mohammad's avatar
Mohammad committed
843
    args = get_args()
844

845
846
847
    (train_dataloader, valid_dataloader, test_dataloader) = (None, None, None)

    print_rank_0('> building train, validation, and test datasets ...')
848
849
850

    # Backward compatibility, assume fixed batch size.
    if args.iteration > 0 and args.consumed_train_samples == 0:
851
852
        assert args.train_samples is None, \
            'only backward compatiblity support for iteration-based training'
mohammad's avatar
mohammad committed
853
        args.consumed_train_samples = args.iteration * args.global_batch_size
854
    if args.iteration > 0 and args.consumed_valid_samples == 0:
855
856
857
        if args.train_samples is None:
            args.consumed_valid_samples = (args.iteration // args.eval_interval) * \
                args.eval_iters * args.global_batch_size
858

859
    # Data loader only on rank 0 of each model parallel group.
860
    if mpu.get_tensor_model_parallel_rank() == 0:
861
862

        # Number of train/valid/test samples.
863
864
865
866
867
868
        if args.train_samples:
            train_samples = args.train_samples
        else:
            train_samples = args.train_iters * args.global_batch_size
        eval_iters = (args.train_iters // args.eval_interval + 1) * \
                     args.eval_iters
869
        test_iters = args.eval_iters
870
        train_val_test_num_samples = [train_samples,
mohammad's avatar
mohammad committed
871
872
                                      eval_iters * args.global_batch_size,
                                      test_iters * args.global_batch_size]
873
874
875
876
877
878
879
880
881
882
        print_rank_0(' > datasets target sizes (minimum size):')
        print_rank_0('    train:      {}'.format(train_val_test_num_samples[0]))
        print_rank_0('    validation: {}'.format(train_val_test_num_samples[1]))
        print_rank_0('    test:       {}'.format(train_val_test_num_samples[2]))

        # Build the datasets.
        train_ds, valid_ds, test_ds = build_train_valid_test_datasets_provider(
            train_val_test_num_samples)

        # Build dataloders.
883
884
885
886
887
        train_dataloader = build_pretraining_data_loader(
            train_ds, args.consumed_train_samples)
        valid_dataloader = build_pretraining_data_loader(
            valid_ds, args.consumed_valid_samples)
        test_dataloader = build_pretraining_data_loader(test_ds, 0)
888
889
890
891
892
893
894
895
896
897
898
899
900

        # Flags to know if we need to do training/validation/testing.
        do_train = train_dataloader is not None and args.train_iters > 0
        do_valid = valid_dataloader is not None and args.eval_iters > 0
        do_test = test_dataloader is not None and args.eval_iters > 0
        # Need to broadcast num_tokens and num_type_tokens.
        flags = torch.cuda.LongTensor(
            [int(do_train), int(do_valid), int(do_test)])
    else:
        flags = torch.cuda.LongTensor([0, 0, 0])

    # Broadcast num tokens.
    torch.distributed.broadcast(flags,
901
902
                                mpu.get_tensor_model_parallel_src_rank(),
                                group=mpu.get_tensor_model_parallel_group())
903
904
905
906
    args.do_train = flags[0].item()
    args.do_valid = flags[1].item()
    args.do_test = flags[2].item()

Vijay Korthikanti's avatar
Vijay Korthikanti committed
907

908
    # Build iterators.
Vijay Korthikanti's avatar
Vijay Korthikanti committed
909
910
911
    dl_type = args.dataloader_type
    assert dl_type in ['single', 'cyclic']

912
    if train_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
913
914
        train_data_iterator = iter(train_dataloader) if dl_type == 'single' \
                              else iter(cyclic_iter(train_dataloader))
915
916
917
    else:
        train_data_iterator = None

918
    if valid_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
919
920
        valid_data_iterator = iter(valid_dataloader) if dl_type == 'single' \
                              else iter(cyclic_iter(valid_dataloader))
921
    else:
922
        valid_data_iterator = None
923

924
    if test_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
925
926
        test_data_iterator = iter(test_dataloader) if dl_type == 'single' \
                             else iter(cyclic_iter(test_dataloader))
927
928
929
    else:
        test_data_iterator = None

930
    return train_data_iterator, valid_data_iterator, test_data_iterator