training.py 36.8 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
Mohammad's avatar
Mohammad committed
29
30
from megatron import get_timers
from megatron import get_tensorboard_writer
31
from megatron import get_current_global_batch_size
mohammad's avatar
mohammad committed
32
from megatron import get_num_microbatches
mohammad's avatar
mohammad committed
33
from megatron import is_last_rank
mohammad's avatar
mohammad committed
34
from megatron import update_num_microbatches
35
from megatron import mpu
Neel Kant's avatar
Neel Kant committed
36
from megatron import print_rank_0
37
from megatron import print_rank_last
Mohammad's avatar
Mohammad committed
38
39
from megatron.checkpointing import load_checkpoint
from megatron.checkpointing import save_checkpoint
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
40
from megatron.model import Float16Module
41
from megatron.model import ModelType
mohammad's avatar
mohammad committed
42
from megatron.optimizer import get_megatron_optimizer
Mohammad's avatar
Mohammad committed
43
from megatron.initialize import initialize_megatron
44
from megatron.initialize import write_args_to_tensorboard
45
46
47
from megatron.learning_rates import AnnealingLR
from megatron.model import DistributedDataParallel as LocalDDP
from megatron.utils import check_adlr_autoresume_termination
48
from megatron.utils import unwrap_model
Vijay Korthikanti's avatar
Vijay Korthikanti committed
49
from megatron.data.data_samplers import build_pretraining_data_loader
mohammad's avatar
mohammad committed
50
from megatron.utils import calc_params_l2_norm
51
from megatron.schedules import get_forward_backward_func
52
from megatron.utils import report_memory
53
54


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
55

56
57
58
59
60
61
62
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))


63
def pretrain(train_valid_test_dataset_provider,
64
             model_provider,
65
             model_type,
66
67
             forward_step_func,
             extra_args_provider=None,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
68
             args_defaults={}):
69
70
71
    """Main training program.

    This function will run the followings in the order provided:
Mohammad's avatar
Mohammad committed
72
73
        1) initialize Megatron.
        2) setup model, optimizer and lr schedule using the model_provider.
74
        3) call train_val_test_data_provider to get train/val/test datasets.
Mohammad's avatar
Mohammad committed
75
        4) train the modle using the forward_step_func.
76
77

    Arguments:
78
79
80
        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
81
            model. By vanilla we mean a simple model on cpu with no fp16 or ddp.
82
        model_type: an enum that specifies the type of model being trained.
Mohammad's avatar
Mohammad committed
83
84
85
86
87
88
89
90
91
        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.
92
93
    """

94
    # Initalize and get arguments, timers, and Tensorboard writer.
95
96
    initialize_megatron(extra_args_provider=extra_args_provider,
                        args_defaults=args_defaults)
97

98
99
100
101
    # 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
102
    start_time_tensor = torch.cuda.DoubleTensor([_TRAIN_START_TIME])
103
104
105
    torch.distributed.all_reduce(start_time_tensor,
                                 op=torch.distributed.ReduceOp.MIN)
    _TRAIN_START_TIME = start_time_tensor.item()
mshoeybi's avatar
mshoeybi committed
106
    print_rank_0('time to initialize megatron (seconds): {:.3f}'.format(
107
108
109
        time.time() - _TRAIN_START_TIME))
    print_datetime('after megatron is initialized')

110
    args = get_args()
Mohammad's avatar
Mohammad committed
111
    timers = get_timers()
112
113

    # Model, optimizer, and learning rate.
114
    timers('model-and-optimizer-setup').start()
115
116
    model, optimizer, lr_scheduler = setup_model_and_optimizer(model_provider,
                                                               model_type)
117
    timers('model-and-optimizer-setup').stop()
118
119
    print_datetime('after model, optimizer, and learning rate '
                   'scheduler are built')
120
121

    # Data stuff.
122
123
    timers('train/valid/test-data-iterators-setup').start()
    if args.virtual_pipeline_model_parallel_size is not None:
124
        all_data_iterators = [
125
126
127
            build_train_valid_test_data_iterators(train_valid_test_dataset_provider)
            for _ in range(len(model))
        ]
128
129
130
        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]
131
132
133
134
135
    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
136
    print_datetime('after dataloaders are built')
Mohammad's avatar
Mohammad committed
137
138

    # Print setup timing.
139
140
    print_rank_0('done with setup ...')
    timers.log(['model-and-optimizer-setup', 'train/valid/test-data-iterators-setup'])
Mohammad's avatar
Mohammad committed
141
    print_rank_0('training ...')
142
143

    iteration = 0
144
    if args.do_train and args.train_iters > 0:
mohammad's avatar
mohammad committed
145
146
147
        iteration = train(forward_step_func,
                          model, optimizer, lr_scheduler,
                          train_data_iterator, valid_data_iterator)
148
    print_datetime('after training is done')
Mohammad's avatar
Mohammad committed
149

150
151
152
    if args.do_valid:
        prefix = 'the end of training for val data'
        evaluate_and_print_results(prefix, forward_step_func,
153
                                   valid_data_iterator, model,
Mohammad's avatar
Mohammad committed
154
                                   iteration, False)
155
156

    if args.save and iteration != 0:
157
        save_checkpoint(iteration, model, optimizer, lr_scheduler)
158
159
160
161
162
163

    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
164
                                   0, True)
165

166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
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]):
182
183
            update_num_microbatches(consumed_samples, consistency_check=False)
            consumed_samples += get_current_global_batch_size()
184
185
            iterations += 1
        # Reset
186
        update_num_microbatches(0, consistency_check=False)
187
188
189
190
191
192
193
194
        # 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))

195

196
def get_model(model_provider_func, model_type):
197
    """Build the model."""
Mohammad's avatar
Mohammad committed
198
    args = get_args()
199
    args.model_type = model_type
200

201
    # Build model.
202
203
    if mpu.get_pipeline_model_parallel_world_size() > 1 and \
       args.virtual_pipeline_model_parallel_size is not None:
204
205
        assert model_type != ModelType.encoder_and_decoder, \
            "Interleaved schedule not supported for model with both encoder and decoder"
206
207
208
        model = []
        for i in range(args.virtual_pipeline_model_parallel_size):
            mpu.set_virtual_pipeline_model_parallel_rank(i)
209
210
211
            # 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()
212
            this_model = model_provider_func(
213
214
215
                pre_process=pre_process,
                post_process=post_process
            )
216
            this_model.model_type = model_type
217
            model.append(this_model)
218
    else:
219
220
        pre_process = mpu.is_pipeline_first_stage()
        post_process = mpu.is_pipeline_last_stage()
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
        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
246

247
248
    if not isinstance(model, list):
        model = [model]
249

250
    # Set tensor model parallel attributes if not set.
mohammad's avatar
mohammad committed
251
252
253
    # 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.
254
255
256
    for model_module in model:
        for param in model_module.parameters():
            mpu.set_defaults_if_not_set_tensor_model_parallel_attributes(param)
257

258
259
    # Print number of parameters.
    if mpu.get_data_parallel_rank() == 0:
260
        print(' > number of parameters on (tensor, pipeline) '
261
              'model parallel rank ({}, {}): {}'.format(
262
263
            mpu.get_tensor_model_parallel_rank(),
            mpu.get_pipeline_model_parallel_rank(),
264
265
            sum([sum([p.nelement() for p in model_module.parameters()])
                 for model_module in model])), flush=True)
266
267

    # GPU allocation.
268
269
    for model_module in model:
        model_module.cuda(torch.cuda.current_device())
270
271

    # Fp16 conversion.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
272
273
    if args.fp16 or args.bf16:
        model = [Float16Module(model_module, args) for model_module in model]
274
275
276

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

282
    if args.DDP_impl == 'local':
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
283
284
285
286
        model = [LocalDDP(model_module,
                          args.accumulate_allreduce_grads_in_fp32,
                          args.use_contiguous_buffers_in_ddp)
                 for model_module in model]
287
288
        return model

289
    raise NotImplementedError('Unknown DDP implementation specified: {}. '
290
                              'Exiting.'.format(args.DDP_impl))
291
292


Mohammad's avatar
Mohammad committed
293
def get_learning_rate_scheduler(optimizer):
294
    """Build the learning rate scheduler."""
Mohammad's avatar
Mohammad committed
295
    args = get_args()
296

297
298
299
300
301
    # 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
302
303
        if args.lr_warmup_fraction is not None:
            warmup_steps = args.lr_warmup_fraction * decay_steps
304
305
        else:
            warmup_steps = args.lr_warmup_iters * args.global_batch_size
306
307
308
309
310
    # 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.
311
        update_train_iters(args)
312
313
314
        if args.lr_decay_samples is None:
            args.lr_decay_samples = args.train_samples
        decay_steps = args.lr_decay_samples
315
316
        if args.lr_warmup_fraction is not None:
            warmup_steps = args.lr_warmup_fraction * decay_steps
317
318
        else:
            warmup_steps = args.lr_warmup_samples
319
    else:
320
321
322
        raise Exception(
            'either train-iters or train-samples should be provided.')

323
324
    lr_scheduler = AnnealingLR(
        optimizer,
325
        max_lr=args.lr,
326
        min_lr=args.min_lr,
327
328
        warmup_steps=warmup_steps,
        decay_steps=decay_steps,
329
        decay_style=args.lr_decay_style,
330
331
332
333
334
335
        use_checkpoint_lr_scheduler=args.use_checkpoint_lr_scheduler,
        override_lr_scheduler=args.override_lr_scheduler)

    return lr_scheduler


336
def setup_model_and_optimizer(model_provider_func, model_type):
337
    """Setup model and optimizer."""
Mohammad's avatar
Mohammad committed
338
    args = get_args()
339

340
    model = get_model(model_provider_func, model_type)
341

342
    unwrapped_model = unwrap_model(model,
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
343
                                   (torchDDP, LocalDDP, Float16Module))
344
345
    optimizer = get_megatron_optimizer(unwrapped_model)

Mohammad's avatar
Mohammad committed
346
    lr_scheduler = get_learning_rate_scheduler(optimizer)
347
348

    if args.load is not None:
349
350
351
352
        timers = get_timers()
        # Extra barrier is added to make sure all ranks report the
        # max time.
        torch.distributed.barrier()
353
        timers('load-checkpoint').start()
354
        args.iteration = load_checkpoint(model, optimizer, lr_scheduler)
355
        torch.distributed.barrier()
356
357
        timers('load-checkpoint').stop()
        timers.log(['load-checkpoint'])
358
359
360
    else:
        args.iteration = 0

mohammad's avatar
mohammad committed
361
    # We only support local DDP with multiple micro-batches.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
362
    if len(model) > 1 or mpu.get_pipeline_model_parallel_world_size() > 1:
mohammad's avatar
mohammad committed
363
364
        assert args.DDP_impl == 'local'

Neel Kant's avatar
Neel Kant committed
365
    # get model without FP16 and/or TorchDDP wrappers
Mostofa Patwary's avatar
Mostofa Patwary committed
366
367
    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
368
        print_rank_0("Initializing ICT from pretrained BERT model")
Mostofa Patwary's avatar
Mostofa Patwary committed
369
        unwrapped_model[0].init_state_dict_from_bert()
Mostofa Patwary's avatar
Mostofa Patwary committed
370
371
        if args.fp16:
            optimizer.reload_model_params()
Neel Kant's avatar
Neel Kant committed
372

373
374
375
    return model, optimizer, lr_scheduler


376
377
378
379
380
381
382
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.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
383
384
385
386
387
    if args.DDP_impl == 'local' and args.use_contiguous_buffers_in_ddp:
        for partition in model:
            partition.zero_grad_buffer()
    else:
        optimizer.zero_grad()
388

389
    forward_backward_func = get_forward_backward_func()
390
391
392
    losses_reduced = forward_backward_func(
        forward_step_func, data_iterator, model,
        optimizer, timers, forward_only=False)
393

394
    # Empty unused memory
395
    if args.empty_unused_memory_each_iter >= 1:
396
397
        torch.cuda.empty_cache()

398
399
    # All-reduce if needed.
    if args.DDP_impl == 'local':
400
        timers('backward-params-all-reduce').start()
401
        for model_module in model:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
402
            model_module.allreduce_gradients()
403
        timers('backward-params-all-reduce').stop()
404

405
406
407
408
    # 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).
409
    timers('backward-embedding-all-reduce').start()
410
    if mpu.is_rank_in_embedding_group(ignore_virtual=True) and \
411
            mpu.get_pipeline_model_parallel_world_size() > 1:
412
413
414
415
        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]
416
417
        else:  # We do not support the interleaved schedule for T5 yet.
            unwrapped_model = model[0]
418
        unwrapped_model = unwrap_model(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
419
            unwrapped_model, (torchDDP, LocalDDP, Float16Module))
420

421
422
        if unwrapped_model.share_word_embeddings:
            word_embeddings_weight = unwrapped_model.word_embeddings_weight()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
423
424
425
426
427
            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())
428
    timers('backward-embedding-all-reduce').stop()
429

430
431
    # Update parameters.
    timers('optimizer').start()
432
    update_successful, grad_norm, num_zeros_in_grad = optimizer.step()
433
434
435
    timers('optimizer').stop()

    # Update learning rate.
436
    if update_successful:
437
438
439
440
        increment = get_num_microbatches() * \
                    args.micro_batch_size * \
                    args.data_parallel_size
        lr_scheduler.step(increment=increment)
mohammad's avatar
mohammad committed
441
        skipped_iter = 0
442
443
444
    else:
        skipped_iter = 1

445
    # Empty unused memory
446
    if args.empty_unused_memory_each_iter >= 2:
447
448
        torch.cuda.empty_cache()

449
    if mpu.is_pipeline_last_stage(ignore_virtual=True):
450
451
452
453
        # Average loss across microbatches.
        loss_reduced = {}
        for key in losses_reduced[0]:
            losses_reduced_for_key = [x[key] for x in losses_reduced]
454
            loss_reduced[key] = sum(losses_reduced_for_key) / len(losses_reduced_for_key)
455
456
        return loss_reduced, skipped_iter, grad_norm, num_zeros_in_grad
    return {}, skipped_iter, grad_norm, num_zeros_in_grad
457
458


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
459
def training_log(loss_dict, total_loss_dict, learning_rate, iteration,
mohammad's avatar
mohammad committed
460
                 loss_scale, report_memory_flag, skipped_iter,
461
                 grad_norm, params_norm, num_zeros_in_grad):
Mohammad's avatar
Mohammad committed
462
463
464
465
    """Log training information such as losses, timing, ...."""
    args = get_args()
    timers = get_timers()
    writer = get_tensorboard_writer()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
466

mohammad's avatar
mohammad committed
467
468
    # Advanced, skipped, and Nan iterations.
    advanced_iters_key = 'advanced iterations'
mohammad's avatar
mohammad committed
469
    skipped_iters_key = 'skipped iterations'
mohammad's avatar
mohammad committed
470
471
472
473
474
475
476
477
478
    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
479
480
    total_loss_dict[skipped_iters_key] = total_loss_dict.get(
        skipped_iters_key, 0) + skipped_iter
mohammad's avatar
mohammad committed
481
    # Update losses and set nan iterations
mohammad's avatar
mohammad committed
482
    got_nan = False
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
483
    for key in loss_dict:
mohammad's avatar
mohammad committed
484
        if not skipped_iter:
485
486
            total_loss_dict[key] = total_loss_dict.get(
                key, torch.cuda.FloatTensor([0.0])) + loss_dict[key]
mohammad's avatar
mohammad committed
487
488
489
490
491
        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
492
            got_nan = got_nan or is_nan
mohammad's avatar
mohammad committed
493
494
    total_loss_dict[nan_iters_key] = total_loss_dict.get(
        nan_iters_key, 0) + int(got_nan)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
495
496
497

    # Logging.
    timers_to_log = []
Neel Kant's avatar
Neel Kant committed
498

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
499
500
501
    def add_to_logging(name):
        if name in timers.timers:
            timers_to_log.append(name)
502
503
504
    add_to_logging('forward-compute')
    add_to_logging('forward-recv')
    add_to_logging('forward-send')
505
    add_to_logging('forward-backward-send-forward-backward-recv')
506
507
508
    add_to_logging('backward-compute')
    add_to_logging('backward-recv')
    add_to_logging('backward-send')
Deepak Narayanan's avatar
Deepak Narayanan committed
509
    add_to_logging('backward-send-forward-recv')
510
    add_to_logging('backward-send-backward-recv')
511
    add_to_logging('backward-params-all-reduce')
512
    add_to_logging('backward-embedding-all-reduce')
513
    add_to_logging('optimizer-copy-to-main-grad')
mohammad's avatar
mohammad committed
514
    add_to_logging('optimizer-unscale-and-check-inf')
515
516
    add_to_logging('optimizer-clip-main-grad')
    add_to_logging('optimizer-copy-main-to-model-params')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
517
    add_to_logging('optimizer')
mohammad's avatar
mohammad committed
518
    add_to_logging('batch-generator')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
519

mohammad's avatar
mohammad committed
520
    # Calculate batch size.
mshoeybi's avatar
mshoeybi committed
521
522
523
    batch_size = args.micro_batch_size * args.data_parallel_size * \
        get_num_microbatches()

mohammad's avatar
mohammad committed
524
525
526
    total_iterations = total_loss_dict[advanced_iters_key] + \
                       total_loss_dict[skipped_iters_key]

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
527
    # Tensorboard values.
528
529
530
531
532
533
534
535
536
537
    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
538
        for key in loss_dict:
mohammad's avatar
mohammad committed
539
540
            writer.add_scalar(key , loss_dict[key], iteration)
            writer.add_scalar(key + ' vs samples', loss_dict[key],
541
                              args.consumed_train_samples)
542
543
544
545
        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)
546
547
548
549
        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)
550
551
552
        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
553
                              args.consumed_train_samples)
mohammad's avatar
mohammad committed
554
555
556
557
        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)
558
559
560
        if args.log_timers_to_tensorboard:
            timers.write(timers_to_log, writer, iteration,
                         normalizer=total_iterations)
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
        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
578
579

    if iteration % args.log_interval == 0:
580
        elapsed_time = timers('interval-time').elapsed()
mohammad's avatar
mohammad committed
581
        elapsed_time_per_iteration = elapsed_time / total_iterations
mshoeybi's avatar
mshoeybi committed
582
        if writer:
583
584
585
            if args.log_timers_to_tensorboard:
                writer.add_scalar('iteration-time',
                                  elapsed_time_per_iteration, iteration)
586
587
        log_string = ' iteration {:8d}/{:8d} |'.format(
            iteration, args.train_iters)
mshoeybi's avatar
mshoeybi committed
588
        log_string += ' consumed samples: {:12d} |'.format(
589
            args.consumed_train_samples)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
590
        log_string += ' elapsed time per iteration (ms): {:.1f} |'.format(
mohammad's avatar
mohammad committed
591
            elapsed_time_per_iteration * 1000.0)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
592
        log_string += ' learning rate: {:.3E} |'.format(learning_rate)
mohammad's avatar
mohammad committed
593
        log_string += ' global batch size: {:5d} |'.format(batch_size)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
594
        for key in total_loss_dict:
mohammad's avatar
mohammad committed
595
596
597
598
            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]))
599
600
601
                if avg > 0.0:
                    log_string += ' {}: {:.6E} |'.format(key, avg)
                total_loss_dict[key] = torch.cuda.FloatTensor([0.0])
602
        log_string += ' loss scale: {:.1f} |'.format(loss_scale)
603
604
        if grad_norm is not None:
            log_string += ' grad norm: {:.3f} |'.format(grad_norm)
605
606
        if num_zeros_in_grad is not None:
            log_string += ' num zeros: {:.1f} |'.format(num_zeros_in_grad)
mohammad's avatar
mohammad committed
607
608
        if params_norm is not None:
            log_string += ' params norm: {:.3f} |'.format(params_norm)
mohammad's avatar
mohammad committed
609
610
        log_string += ' number of skipped iterations: {:3d} |'.format(
            total_loss_dict[skipped_iters_key])
mohammad's avatar
mohammad committed
611
        log_string += ' number of nan iterations: {:3d} |'.format(
mohammad's avatar
mohammad committed
612
613
            total_loss_dict[nan_iters_key])
        total_loss_dict[advanced_iters_key] = 0
mohammad's avatar
mohammad committed
614
        total_loss_dict[skipped_iters_key] = 0
mohammad's avatar
mohammad committed
615
        total_loss_dict[nan_iters_key] = 0
616
        print_rank_last(log_string)
617
618
619
        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
620
621
622
623
624
625
            report_memory_flag = False
        timers.log(timers_to_log, normalizer=args.log_interval)

    return report_memory_flag


626
627
628
629
630
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()
631
    timers('save-checkpoint').start()
632
633
    save_checkpoint(iteration, model, optimizer, lr_scheduler)
    torch.distributed.barrier()
634
635
    timers('save-checkpoint').stop()
    timers.log(['save-checkpoint'])
636
637


638
def train(forward_step_func, model, optimizer, lr_scheduler,
639
          train_data_iterator, valid_data_iterator):
640
    """Train the model function."""
Mohammad's avatar
Mohammad committed
641
642
    args = get_args()
    timers = get_timers()
643

644
645
646
    # Write args to tensorboard
    write_args_to_tensorboard()

647
    # Turn on training mode which enables dropout.
648
649
    for model_module in model:
        model_module.train()
650
651
652
653
654
655
656

    # Tracking loss.
    total_loss_dict = {}

    # Iterations.
    iteration = args.iteration

657
    timers('interval-time').start()
658
    print_datetime('before the start of training step')
659
660
    report_memory_flag = True
    while iteration < args.train_iters:
mohammad's avatar
mohammad committed
661
        update_num_microbatches(args.consumed_train_samples)
662
663
664
665
666
667
        loss_dict, skipped_iter, grad_norm, num_zeros_in_grad = \
            train_step(forward_step_func,
                       train_data_iterator,
                       model,
                       optimizer,
                       lr_scheduler)
668
        iteration += 1
669
        args.consumed_train_samples += mpu.get_data_parallel_world_size() * \
670
                                       args.micro_batch_size * \
mohammad's avatar
mohammad committed
671
                                       get_num_microbatches()
672
673

        # Logging.
674
        loss_scale = optimizer.get_loss_scale().item()
675
676
677
        params_norm = None
        if args.log_params_norm:
            params_norm = calc_params_l2_norm(model)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
678
679
        report_memory_flag = training_log(loss_dict, total_loss_dict,
                                          optimizer.param_groups[0]['lr'],
Mohammad's avatar
Mohammad committed
680
                                          iteration, loss_scale,
681
                                          report_memory_flag, skipped_iter,
682
                                          grad_norm, params_norm, num_zeros_in_grad)
683
684

        # Autoresume
685
686
        if args.adlr_autoresume and \
           (iteration % args.adlr_autoresume_interval == 0):
687
            check_adlr_autoresume_termination(iteration, model, optimizer,
688
                                              lr_scheduler)
689
690
691
692
693
694

        # 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,
695
                                       valid_data_iterator, model,
Mohammad's avatar
Mohammad committed
696
                                       iteration, False)
697

698
699
700
701
702
703
704
705
        # Checkpointing
        saved_checkpoint = False
        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

706
707
708
709
710
711
712
713
714
715
716
717
        # 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)
718
                print_datetime('exiting program after {} minutes'.format(train_time))
719
720
                sys.exit()

721
        # Exiting based on iterations
722
        if args.exit_interval and iteration % args.exit_interval == 0:
723
724
725
            if not saved_checkpoint:
                save_checkpoint_and_time(iteration, model, optimizer,
                                         lr_scheduler)
726
            torch.distributed.barrier()
727
            print_datetime('exiting program at iteration {}'.format(iteration))
Mohammad's avatar
Mohammad committed
728
            sys.exit()
729

730

mohammad's avatar
mohammad committed
731
    return iteration
732
733


Mohammad's avatar
Mohammad committed
734
def evaluate(forward_step_func, data_iterator, model, verbose=False):
735
    """Evaluation."""
Mohammad's avatar
Mohammad committed
736
    args = get_args()
737
738

    # Turn on evaluation mode which disables dropout.
739
740
    for model_module in model:
        model_module.eval()
741
742
743
744
745
746
747
748
749
750

    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))
751

752
            forward_backward_func = get_forward_backward_func()
753
754
755
756
            loss_dicts = forward_backward_func(
                forward_step_func, data_iterator, model, optimizer=None,
                timers=None, forward_only=True)

757
758
759
760
            # Empty unused memory
            if args.empty_unused_memory_each_iter >= 1:
                torch.cuda.empty_cache()

761
762
763
            if mpu.is_pipeline_last_stage(ignore_virtual=True):
                # Reduce across processes.
                for loss_dict in loss_dicts:
764
                    for key in loss_dict:
765
766
                        total_loss_dict[key] = total_loss_dict.get(
                            key, torch.cuda.FloatTensor([0.0])) + loss_dict[key]
767

768
            args.consumed_valid_samples += mpu.get_data_parallel_world_size() \
769
                                           * args.micro_batch_size \
mohammad's avatar
mohammad committed
770
                                           * get_num_microbatches()
771
    # Move model back to the train mode.
772
773
    for model_module in model:
        model_module.train()
774
775

    for key in total_loss_dict:
mohammad's avatar
mohammad committed
776
        total_loss_dict[key] /= args.eval_iters * get_num_microbatches()
777
778
779
780
781

    return total_loss_dict

def evaluate_and_print_results(prefix, forward_step_func,
                               data_iterator, model,
Mohammad's avatar
Mohammad committed
782
                               iteration, verbose=False):
783
    """Helper function to evaluate and dump results on screen."""
784
    args = get_args()
Mohammad's avatar
Mohammad committed
785
786
787
    writer = get_tensorboard_writer()

    total_loss_dict = evaluate(forward_step_func, data_iterator, model, verbose)
788
789
790
791
792
    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
793
        if writer:
mohammad's avatar
mohammad committed
794
            writer.add_scalar('{} validation'.format(key),
795
796
                              total_loss_dict[key].item(),
                              iteration)
mohammad's avatar
mohammad committed
797
            writer.add_scalar('{} validation vs samples'.format(key),
798
799
                              total_loss_dict[key].item(),
                              args.consumed_train_samples)
800
            if args.log_validation_ppl_to_tensorboard:
mohammad's avatar
mohammad committed
801
                writer.add_scalar('{} validation ppl'.format(key), ppl,
802
                                  iteration)
mohammad's avatar
mohammad committed
803
                writer.add_scalar('{} validation ppl vs samples'.format(key),
804
                                  ppl, args.consumed_train_samples)
805
806

    length = len(string) + 1
807
808
809
    print_rank_last('-' * length)
    print_rank_last(string)
    print_rank_last('-' * length)
810
811


Vijay Korthikanti's avatar
Vijay Korthikanti committed
812
def cyclic_iter(iter):
813
    while True:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
814
        for x in iter:
815
816
            yield x

817
818
819
def build_train_valid_test_data_iterators(
        build_train_valid_test_datasets_provider):
    """XXX"""
Mohammad's avatar
Mohammad committed
820
    args = get_args()
821

822
823
824
    (train_dataloader, valid_dataloader, test_dataloader) = (None, None, None)

    print_rank_0('> building train, validation, and test datasets ...')
825
826
827

    # Backward compatibility, assume fixed batch size.
    if args.iteration > 0 and args.consumed_train_samples == 0:
828
829
        assert args.train_samples is None, \
            'only backward compatiblity support for iteration-based training'
mohammad's avatar
mohammad committed
830
        args.consumed_train_samples = args.iteration * args.global_batch_size
831
    if args.iteration > 0 and args.consumed_valid_samples == 0:
832
833
834
        if args.train_samples is None:
            args.consumed_valid_samples = (args.iteration // args.eval_interval) * \
                args.eval_iters * args.global_batch_size
835

836
    # Data loader only on rank 0 of each model parallel group.
837
    if mpu.get_tensor_model_parallel_rank() == 0:
838
839

        # Number of train/valid/test samples.
840
841
842
843
844
845
        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
846
        test_iters = args.eval_iters
847
        train_val_test_num_samples = [train_samples,
mohammad's avatar
mohammad committed
848
849
                                      eval_iters * args.global_batch_size,
                                      test_iters * args.global_batch_size]
850
851
852
853
854
855
856
857
858
859
        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.
860
861
862
863
864
        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)
865
866
867
868
869
870
871
872
873
874
875
876
877

        # 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,
878
879
                                mpu.get_tensor_model_parallel_src_rank(),
                                group=mpu.get_tensor_model_parallel_group())
880
881
882
883
    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
884

885
    # Build iterators.
Vijay Korthikanti's avatar
Vijay Korthikanti committed
886
887
888
    dl_type = args.dataloader_type
    assert dl_type in ['single', 'cyclic']

889
    if train_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
890
891
        train_data_iterator = iter(train_dataloader) if dl_type == 'single' \
                              else iter(cyclic_iter(train_dataloader))
892
893
894
    else:
        train_data_iterator = None

895
    if valid_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
896
897
        valid_data_iterator = iter(valid_dataloader) if dl_type == 'single' \
                              else iter(cyclic_iter(valid_dataloader))
898
    else:
899
        valid_data_iterator = None
900

901
    if test_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
902
903
        test_data_iterator = iter(test_dataloader) if dl_type == 'single' \
                             else iter(cyclic_iter(test_dataloader))
904
905
906
    else:
        test_data_iterator = None

907
    return train_data_iterator, valid_data_iterator, test_data_iterator