training.py 42.4 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
import time
# The earliest we can measure the start time.
_TRAIN_START_TIME = time.time()
24
25
import torch
from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
26
from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
27

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
from megatron.optimizer_param_scheduler import OptimizerParamScheduler
47
48
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
from megatron.model.vision.knn_monitor import compute_feature_bank
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
             forward_step_func,
68
             process_non_loss_data_func=None,
69
             extra_args_provider=None,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
70
             args_defaults={}):
71
72
73
    """Main training program.

    This function will run the followings in the order provided:
Mohammad's avatar
Mohammad committed
74
75
        1) initialize Megatron.
        2) setup model, optimizer and lr schedule using the model_provider.
76
        3) call train_val_test_data_provider to get train/val/test datasets.
Mohammad's avatar
Mohammad committed
77
        4) train the modle using the forward_step_func.
78
79

    Arguments:
80
81
82
        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
83
            model. By vanilla we mean a simple model on cpu with no fp16 or ddp.
84
        model_type: an enum that specifies the type of model being trained.
Mohammad's avatar
Mohammad committed
85
86
87
88
89
        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.
Vijay Korthikanti's avatar
Vijay Korthikanti committed
90
91
92
93
        process_non_loss_data_func: a function to post process outputs of the
            network. It can be used for dumping output tensors (e.g images) to
            tensorboard. It takes `collected data`(list of tensors),
            `current iteration index` and `tensorboard writer` as arguments.
Mohammad's avatar
Mohammad committed
94
95
96
97
        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.
98
99
    """

100
    # Initalize and get arguments, timers, and Tensorboard writer.
101
102
    initialize_megatron(extra_args_provider=extra_args_provider,
                        args_defaults=args_defaults)
103

104
105
106
107
    # 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
108
    start_time_tensor = torch.cuda.DoubleTensor([_TRAIN_START_TIME])
109
110
111
    torch.distributed.all_reduce(start_time_tensor,
                                 op=torch.distributed.ReduceOp.MIN)
    _TRAIN_START_TIME = start_time_tensor.item()
mshoeybi's avatar
mshoeybi committed
112
    print_rank_0('time to initialize megatron (seconds): {:.3f}'.format(
113
114
115
        time.time() - _TRAIN_START_TIME))
    print_datetime('after megatron is initialized')

116
    args = get_args()
Mohammad's avatar
Mohammad committed
117
    timers = get_timers()
118
119

    # Model, optimizer, and learning rate.
120
    timers('model-and-optimizer-setup').start()
121
    model, optimizer, opt_param_scheduler = setup_model_and_optimizer(model_provider,
122
                                                               model_type)
123
    timers('model-and-optimizer-setup').stop()
124
125
    print_datetime('after model, optimizer, and learning rate '
                   'scheduler are built')
126
127

    # Data stuff.
128
129
    timers('train/valid/test-data-iterators-setup').start()
    if args.virtual_pipeline_model_parallel_size is not None:
130
        all_data_iterators = [
131
132
133
            build_train_valid_test_data_iterators(train_valid_test_dataset_provider)
            for _ in range(len(model))
        ]
134
135
136
        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]
137
138
139
140
141
    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
142
    print_datetime('after dataloaders are built')
Mohammad's avatar
Mohammad committed
143
144

    # Print setup timing.
145
146
    print_rank_0('done with setup ...')
    timers.log(['model-and-optimizer-setup', 'train/valid/test-data-iterators-setup'])
Mohammad's avatar
Mohammad committed
147
    print_rank_0('training ...')
148
149

    iteration = 0
150
    if args.do_train and args.train_iters > 0:
mohammad's avatar
mohammad committed
151
        iteration = train(forward_step_func,
152
                          model, optimizer, opt_param_scheduler,
153
154
                          train_data_iterator, valid_data_iterator,
                          process_non_loss_data_func)
155
    print_datetime('after training is done')
Mohammad's avatar
Mohammad committed
156

157
158
159
    if args.do_valid:
        prefix = 'the end of training for val data'
        evaluate_and_print_results(prefix, forward_step_func,
160
                                   valid_data_iterator, model,
161
162
                                   iteration, process_non_loss_data_func,
                                   False)
163
164

    if args.save and iteration != 0:
165
        save_checkpoint(iteration, model, optimizer, opt_param_scheduler)
166
167
168
169
170
171

    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,
172
173
                                   0, process_non_loss_data_func,
                                   True)
174

175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
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]):
191
192
            update_num_microbatches(consumed_samples, consistency_check=False)
            consumed_samples += get_current_global_batch_size()
193
194
            iterations += 1
        # Reset
195
        update_num_microbatches(0, consistency_check=False)
196
197
198
199
200
201
202
203
        # 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))

204

205
def get_model(model_provider_func, model_type=ModelType.encoder_or_decoder, wrap_with_ddp=True):
206
    """Build the model."""
Mohammad's avatar
Mohammad committed
207
    args = get_args()
208
    args.model_type = model_type
209

210
    # Build model.
211
212
    if mpu.get_pipeline_model_parallel_world_size() > 1 and \
       args.virtual_pipeline_model_parallel_size is not None:
213
214
        assert model_type != ModelType.encoder_and_decoder, \
            "Interleaved schedule not supported for model with both encoder and decoder"
215
216
217
        model = []
        for i in range(args.virtual_pipeline_model_parallel_size):
            mpu.set_virtual_pipeline_model_parallel_rank(i)
218
219
220
            # 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()
221
            this_model = model_provider_func(
222
223
224
                pre_process=pre_process,
                post_process=post_process
            )
225
            this_model.model_type = model_type
226
            model.append(this_model)
227
    else:
228
229
        pre_process = mpu.is_pipeline_first_stage()
        post_process = mpu.is_pipeline_last_stage()
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
        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
255

256
257
    if not isinstance(model, list):
        model = [model]
258

259
    # Set tensor model parallel attributes if not set.
mohammad's avatar
mohammad committed
260
261
262
    # 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.
263
264
265
    for model_module in model:
        for param in model_module.parameters():
            mpu.set_defaults_if_not_set_tensor_model_parallel_attributes(param)
266

267
268
    # Print number of parameters.
    if mpu.get_data_parallel_rank() == 0:
269
        print(' > number of parameters on (tensor, pipeline) '
270
              'model parallel rank ({}, {}): {}'.format(
271
272
            mpu.get_tensor_model_parallel_rank(),
            mpu.get_pipeline_model_parallel_rank(),
273
274
            sum([sum([p.nelement() for p in model_module.parameters()])
                 for model_module in model])), flush=True)
275
276

    # GPU allocation.
277
278
    for model_module in model:
        model_module.cuda(torch.cuda.current_device())
279
280

    # Fp16 conversion.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
281
282
    if args.fp16 or args.bf16:
        model = [Float16Module(model_module, args) for model_module in model]
283

284
285
286
287
288
289
    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]
290

291
292
293
294
295
        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]
296
297
298
299
            # broad cast params from data parallel src rank to other data parallel ranks
            if args.data_parallel_random_init:
                for model_module in model:
                    model_module.broadcast_params()
300
301
302
        else:
            raise NotImplementedError('Unknown DDP implementation specified: '
                                      '{}. Exiting.'.format(args.DDP_impl))
303

304
    return model
305
306


307
def get_optimizer_param_scheduler(optimizer):
308
    """Build the learning rate scheduler."""
Mohammad's avatar
Mohammad committed
309
    args = get_args()
310

311
312
313
314
    # Iteration-based training.
    if args.train_iters:
        if args.lr_decay_iters is None:
            args.lr_decay_iters = args.train_iters
Vijay Korthikanti's avatar
Vijay Korthikanti committed
315
316
        lr_decay_steps = args.lr_decay_iters * args.global_batch_size
        wd_incr_steps = args.train_iters * args.global_batch_size
317
        if args.lr_warmup_fraction is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
318
            lr_warmup_steps = args.lr_warmup_fraction * lr_decay_steps
319
        else:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
320
            lr_warmup_steps = args.lr_warmup_iters * args.global_batch_size
321
322
323
324
325
    # 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.
326
        update_train_iters(args)
327
328
        if args.lr_decay_samples is None:
            args.lr_decay_samples = args.train_samples
Vijay Korthikanti's avatar
Vijay Korthikanti committed
329
330
        lr_decay_steps = args.lr_decay_samples
        wd_incr_steps = args.train_samples
331
        if args.lr_warmup_fraction is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
332
            lr_warmup_steps = args.lr_warmup_fraction * lr_decay_steps
333
        else:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
334
            lr_warmup_steps = args.lr_warmup_samples
335
    else:
336
337
338
        raise Exception(
            'either train-iters or train-samples should be provided.')

339
    opt_param_scheduler = OptimizerParamScheduler(
340
        optimizer,
341
        max_lr=args.lr,
342
        min_lr=args.min_lr,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
343
344
345
        lr_warmup_steps=lr_warmup_steps,
        lr_decay_steps=lr_decay_steps,
        lr_decay_style=args.lr_decay_style,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
346
347
        start_wd=args.start_weight_decay,
        end_wd=args.end_weight_decay,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
348
        wd_incr_steps=wd_incr_steps,
Vijay Korthikanti's avatar
Vijay Korthikanti committed
349
        wd_incr_style=args.weight_decay_incr_style,
350
351
        use_checkpoint_opt_param_scheduler=args.use_checkpoint_opt_param_scheduler,
        override_opt_param_scheduler=args.override_opt_param_scheduler)
352

353
    return opt_param_scheduler
354
355


356
357
358
359
360
def setup_model_and_optimizer(model_provider_func,
                              model_type,
                              no_wd_decay_cond=None,
                              scale_lr_cond=None,
                              lr_mult=1.0):
361
    """Setup model and optimizer."""
Mohammad's avatar
Mohammad committed
362
    args = get_args()
363

364
    model = get_model(model_provider_func, model_type)
365

366
    unwrapped_model = unwrap_model(model,
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
367
                                   (torchDDP, LocalDDP, Float16Module))
368
369
    optimizer = get_megatron_optimizer(unwrapped_model, no_wd_decay_cond,
                                       scale_lr_cond, lr_mult)
370

371
    opt_param_scheduler = get_optimizer_param_scheduler(optimizer)
372
373

    if args.load is not None:
374
375
376
377
        timers = get_timers()
        # Extra barrier is added to make sure all ranks report the
        # max time.
        torch.distributed.barrier()
378
        timers('load-checkpoint').start()
379
        args.iteration = load_checkpoint(model, optimizer, opt_param_scheduler)
380
        torch.distributed.barrier()
381
382
        timers('load-checkpoint').stop()
        timers.log(['load-checkpoint'])
383
384
385
    else:
        args.iteration = 0

mohammad's avatar
mohammad committed
386
    # We only support local DDP with multiple micro-batches.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
387
    if len(model) > 1 or mpu.get_pipeline_model_parallel_world_size() > 1:
mohammad's avatar
mohammad committed
388
389
        assert args.DDP_impl == 'local'

Neel Kant's avatar
Neel Kant committed
390
    # get model without FP16 and/or TorchDDP wrappers
Mostofa Patwary's avatar
Mostofa Patwary committed
391
392
    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
393
        print_rank_0("Initializing ICT from pretrained BERT model")
Mostofa Patwary's avatar
Mostofa Patwary committed
394
        unwrapped_model[0].init_state_dict_from_bert()
Mostofa Patwary's avatar
Mostofa Patwary committed
395
396
        if args.fp16:
            optimizer.reload_model_params()
Neel Kant's avatar
Neel Kant committed
397

398
    return model, optimizer, opt_param_scheduler
399
400


401
def train_step(forward_step_func, data_iterator,
402
               model, optimizer, opt_param_scheduler):
403
404
405
406
407
    """Single training step."""
    args = get_args()
    timers = get_timers()

    # Set grad to zero.
408
    if args.DDP_impl == 'local' and args.use_contiguous_buffers_in_local_ddp:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
409
410
        for partition in model:
            partition.zero_grad_buffer()
411
    optimizer.zero_grad()
412

413
    forward_backward_func = get_forward_backward_func()
414
415
416
    losses_reduced = forward_backward_func(
        forward_step_func, data_iterator, model,
        optimizer, timers, forward_only=False)
417

418
    # Empty unused memory
Lawrence McAfee's avatar
Lawrence McAfee committed
419
    if args.empty_unused_memory_level >= 1:
420
421
        torch.cuda.empty_cache()

422
423
    # All-reduce layernorm parameters across model parallel nodes
    # when sequence parallelism is used
424
    if mpu.get_tensor_model_parallel_world_size() > 1 and \
Vijay Korthikanti's avatar
Vijay Korthikanti committed
425
            args.sequence_parallel:
426
427
428
429
430
        grads = []
        for model_module in model:
            unwrapped_model = unwrap_model( 
                model_module, (torchDDP, LocalDDP, Float16Module))
            for param in unwrapped_model.parameters():
431
432
433
                if getattr(param, 'sequence_parallel', False):
                    grad = param.main_grad if args.DDP_impl == 'local' else param.grad
                    grads.append(grad.data)
434
435
436
437
438
439
440
        coalesced = _flatten_dense_tensors(grads)
        torch.distributed.all_reduce(
            coalesced, group=mpu.get_tensor_model_parallel_group())
        for buf, synced in zip(grads, _unflatten_dense_tensors(
                coalesced, grads)):
            buf.copy_(synced)

441
442
    # All-reduce if needed.
    if args.DDP_impl == 'local':
443
        timers('backward-params-all-reduce').start()
444
        for model_module in model:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
445
            model_module.allreduce_gradients()
446
        timers('backward-params-all-reduce').stop()
447

448
449
450
451
    # 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).
452
    timers('backward-embedding-all-reduce').start()
453
    if mpu.is_rank_in_embedding_group(ignore_virtual=True) and \
454
            mpu.get_pipeline_model_parallel_world_size() > 1:
455
456
457
458
        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]
459
460
        else:  # We do not support the interleaved schedule for T5 yet.
            unwrapped_model = model[0]
461
        unwrapped_model = unwrap_model(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
462
            unwrapped_model, (torchDDP, LocalDDP, Float16Module))
463

464
465
        if unwrapped_model.share_word_embeddings:
            word_embeddings_weight = unwrapped_model.word_embeddings_weight()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
466
467
468
469
470
            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
471

Vijay Korthikanti's avatar
Vijay Korthikanti committed
472
473
474
    # All-reduce position_embeddings grad across first (encoder) and split (decoder) 
    # stages to ensure that position embeddings parameters stay in sync.
    # This should only run for T5 models with pipeline parallelism
Vijay Korthikanti's avatar
Vijay Korthikanti committed
475
476
477
478
479
480
    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))
Vijay Korthikanti's avatar
Vijay Korthikanti committed
481
482
        assert args.DDP_impl == 'local', \
            'T5 model is only supported with local DDP mode'
Vijay Korthikanti's avatar
Vijay Korthikanti committed
483
484
        grad = unwrapped_model.language_model.embedding.position_embeddings.weight.main_grad
        torch.distributed.all_reduce(grad, group=mpu.get_position_embedding_group())
485
    timers('backward-embedding-all-reduce').stop()
486

Vijay Korthikanti's avatar
Vijay Korthikanti committed
487
    if args.vision_pretraining and args.vision_pretraining_type == "dino":
488
489
490
491
492
        unwrapped_model = unwrap_model(model[0],
                                       (torchDDP, LocalDDP, Float16Module))
        unwrapped_model.cancel_gradients_last_layer(args.curr_iteration)


493
494
    # Update parameters.
    timers('optimizer').start()
495
    update_successful, grad_norm, num_zeros_in_grad = optimizer.step()
496
497
    timers('optimizer').stop()

Vijay Korthikanti's avatar
Vijay Korthikanti committed
498
    if args.vision_pretraining and args.vision_pretraining_type == "dino":
499
500
501
502
503
        unwrapped_model = unwrap_model(model[0],
                                       (torchDDP, LocalDDP, Float16Module))
        unwrapped_model.update_momentum(args.curr_iteration)


504
    # Update learning rate.
505
    if update_successful:
506
507
508
        increment = get_num_microbatches() * \
                    args.micro_batch_size * \
                    args.data_parallel_size
509
        opt_param_scheduler.step(increment=increment)
mohammad's avatar
mohammad committed
510
        skipped_iter = 0
511
512
513
    else:
        skipped_iter = 1

514
    # Empty unused memory
Lawrence McAfee's avatar
Lawrence McAfee committed
515
    if args.empty_unused_memory_level >= 2:
516
517
        torch.cuda.empty_cache()

518
    if mpu.is_pipeline_last_stage(ignore_virtual=True):
519
520
521
522
        # Average loss across microbatches.
        loss_reduced = {}
        for key in losses_reduced[0]:
            losses_reduced_for_key = [x[key] for x in losses_reduced]
523
            loss_reduced[key] = sum(losses_reduced_for_key) / len(losses_reduced_for_key)
524
525
        return loss_reduced, skipped_iter, grad_norm, num_zeros_in_grad
    return {}, skipped_iter, grad_norm, num_zeros_in_grad
526
527


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
528
def training_log(loss_dict, total_loss_dict, learning_rate, iteration,
mohammad's avatar
mohammad committed
529
                 loss_scale, report_memory_flag, skipped_iter,
530
                 grad_norm, params_norm, num_zeros_in_grad):
Mohammad's avatar
Mohammad committed
531
532
533
534
    """Log training information such as losses, timing, ...."""
    args = get_args()
    timers = get_timers()
    writer = get_tensorboard_writer()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
535

mohammad's avatar
mohammad committed
536
537
    # Advanced, skipped, and Nan iterations.
    advanced_iters_key = 'advanced iterations'
mohammad's avatar
mohammad committed
538
    skipped_iters_key = 'skipped iterations'
mohammad's avatar
mohammad committed
539
540
541
542
543
544
545
546
547
    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
548
549
    total_loss_dict[skipped_iters_key] = total_loss_dict.get(
        skipped_iters_key, 0) + skipped_iter
mohammad's avatar
mohammad committed
550
    # Update losses and set nan iterations
mohammad's avatar
mohammad committed
551
    got_nan = False
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
552
    for key in loss_dict:
mohammad's avatar
mohammad committed
553
        if not skipped_iter:
554
555
            total_loss_dict[key] = total_loss_dict.get(
                key, torch.cuda.FloatTensor([0.0])) + loss_dict[key]
mohammad's avatar
mohammad committed
556
557
558
559
560
        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
561
            got_nan = got_nan or is_nan
mohammad's avatar
mohammad committed
562
563
    total_loss_dict[nan_iters_key] = total_loss_dict.get(
        nan_iters_key, 0) + int(got_nan)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
564
565
566

    # Logging.
    timers_to_log = []
Neel Kant's avatar
Neel Kant committed
567

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
568
569
570
    def add_to_logging(name):
        if name in timers.timers:
            timers_to_log.append(name)
571
572
573
    add_to_logging('forward-compute')
    add_to_logging('forward-recv')
    add_to_logging('forward-send')
574
    add_to_logging('forward-backward-send-forward-backward-recv')
575
576
577
    add_to_logging('backward-compute')
    add_to_logging('backward-recv')
    add_to_logging('backward-send')
Deepak Narayanan's avatar
Deepak Narayanan committed
578
    add_to_logging('backward-send-forward-recv')
579
    add_to_logging('backward-send-backward-recv')
580
    add_to_logging('backward-params-all-reduce')
581
    add_to_logging('backward-embedding-all-reduce')
582
    add_to_logging('optimizer-copy-to-main-grad')
mohammad's avatar
mohammad committed
583
    add_to_logging('optimizer-unscale-and-check-inf')
584
585
    add_to_logging('optimizer-clip-main-grad')
    add_to_logging('optimizer-copy-main-to-model-params')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
586
    add_to_logging('optimizer')
mohammad's avatar
mohammad committed
587
    add_to_logging('batch-generator')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
588

mohammad's avatar
mohammad committed
589
    # Calculate batch size.
mshoeybi's avatar
mshoeybi committed
590
591
592
    batch_size = args.micro_batch_size * args.data_parallel_size * \
        get_num_microbatches()

mohammad's avatar
mohammad committed
593
594
595
    total_iterations = total_loss_dict[advanced_iters_key] + \
                       total_loss_dict[skipped_iters_key]

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
596
    # Tensorboard values.
597
598
599
600
601
602
603
604
605
606
    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
607
        for key in loss_dict:
mohammad's avatar
mohammad committed
608
609
            writer.add_scalar(key , loss_dict[key], iteration)
            writer.add_scalar(key + ' vs samples', loss_dict[key],
610
                              args.consumed_train_samples)
611
612
613
614
        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)
615
616
617
618
        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)
619
620
621
622
        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)
623
624
625
        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
626
                              args.consumed_train_samples)
mohammad's avatar
mohammad committed
627
628
629
630
        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)
631
632
633
        if args.log_timers_to_tensorboard:
            timers.write(timers_to_log, writer, iteration,
                         normalizer=total_iterations)
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
        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
651
652

    if iteration % args.log_interval == 0:
653
        elapsed_time = timers('interval-time').elapsed()
mohammad's avatar
mohammad committed
654
        elapsed_time_per_iteration = elapsed_time / total_iterations
mshoeybi's avatar
mshoeybi committed
655
        if writer:
656
657
658
            if args.log_timers_to_tensorboard:
                writer.add_scalar('iteration-time',
                                  elapsed_time_per_iteration, iteration)
659
660
        log_string = ' iteration {:8d}/{:8d} |'.format(
            iteration, args.train_iters)
mshoeybi's avatar
mshoeybi committed
661
        log_string += ' consumed samples: {:12d} |'.format(
662
            args.consumed_train_samples)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
663
        log_string += ' elapsed time per iteration (ms): {:.1f} |'.format(
mohammad's avatar
mohammad committed
664
            elapsed_time_per_iteration * 1000.0)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
665
        log_string += ' learning rate: {:.3E} |'.format(learning_rate)
mohammad's avatar
mohammad committed
666
        log_string += ' global batch size: {:5d} |'.format(batch_size)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
667
        for key in total_loss_dict:
mohammad's avatar
mohammad committed
668
669
670
671
            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]))
672
673
674
                if avg > 0.0:
                    log_string += ' {}: {:.6E} |'.format(key, avg)
                total_loss_dict[key] = torch.cuda.FloatTensor([0.0])
675
        log_string += ' loss scale: {:.1f} |'.format(loss_scale)
676
677
        if grad_norm is not None:
            log_string += ' grad norm: {:.3f} |'.format(grad_norm)
678
679
        if num_zeros_in_grad is not None:
            log_string += ' num zeros: {:.1f} |'.format(num_zeros_in_grad)
mohammad's avatar
mohammad committed
680
681
        if params_norm is not None:
            log_string += ' params norm: {:.3f} |'.format(params_norm)
mohammad's avatar
mohammad committed
682
683
        log_string += ' number of skipped iterations: {:3d} |'.format(
            total_loss_dict[skipped_iters_key])
mohammad's avatar
mohammad committed
684
        log_string += ' number of nan iterations: {:3d} |'.format(
mohammad's avatar
mohammad committed
685
686
            total_loss_dict[nan_iters_key])
        total_loss_dict[advanced_iters_key] = 0
mohammad's avatar
mohammad committed
687
        total_loss_dict[skipped_iters_key] = 0
mohammad's avatar
mohammad committed
688
        total_loss_dict[nan_iters_key] = 0
689
        print_rank_last(log_string)
690
691
692
        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
693
694
695
696
697
698
            report_memory_flag = False
        timers.log(timers_to_log, normalizer=args.log_interval)

    return report_memory_flag


699
def save_checkpoint_and_time(iteration, model, optimizer, opt_param_scheduler):
700
701
702
703
    timers = get_timers()
    # Extra barrier is added to make sure
    # all ranks report the max time.
    torch.distributed.barrier()
704
    timers('save-checkpoint').start()
705
    save_checkpoint(iteration, model, optimizer, opt_param_scheduler)
706
    torch.distributed.barrier()
707
708
    timers('save-checkpoint').stop()
    timers.log(['save-checkpoint'])
709
710


711
def train(forward_step_func, model, optimizer, opt_param_scheduler,
712
713
          train_data_iterator, valid_data_iterator,
          process_non_loss_data_func):
714
    """Train the model function."""
Mohammad's avatar
Mohammad committed
715
716
    args = get_args()
    timers = get_timers()
717

718
719
720
    # Write args to tensorboard
    write_args_to_tensorboard()

721
    # Turn on training mode which enables dropout.
722
723
    for model_module in model:
        model_module.train()
724
725
726
727
728
729
730

    # Tracking loss.
    total_loss_dict = {}

    # Iterations.
    iteration = args.iteration

731
    timers('interval-time').start()
732
    print_datetime('before the start of training step')
733
734
    report_memory_flag = True
    while iteration < args.train_iters:
mohammad's avatar
mohammad committed
735
        update_num_microbatches(args.consumed_train_samples)
Vijay Korthikanti's avatar
Vijay Korthikanti committed
736
        args.curr_iteration = iteration
737
738
739
740
741
        loss_dict, skipped_iter, grad_norm, num_zeros_in_grad = \
            train_step(forward_step_func,
                       train_data_iterator,
                       model,
                       optimizer,
742
                       opt_param_scheduler)
743
        iteration += 1
744
        args.consumed_train_samples += mpu.get_data_parallel_world_size() * \
745
                                       args.micro_batch_size * \
mohammad's avatar
mohammad committed
746
                                       get_num_microbatches()
747
748

        # Logging.
749
        loss_scale = optimizer.get_loss_scale().item()
750
751
752
        params_norm = None
        if args.log_params_norm:
            params_norm = calc_params_l2_norm(model)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
753
754
        report_memory_flag = training_log(loss_dict, total_loss_dict,
                                          optimizer.param_groups[0]['lr'],
Mohammad's avatar
Mohammad committed
755
                                          iteration, loss_scale,
756
                                          report_memory_flag, skipped_iter,
757
                                          grad_norm, params_norm, num_zeros_in_grad)
758
759

        # Autoresume
760
761
        if args.adlr_autoresume and \
           (iteration % args.adlr_autoresume_interval == 0):
762
            check_adlr_autoresume_termination(iteration, model, optimizer,
763
                                              opt_param_scheduler)
764
765
766
767
768
769

        # 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,
770
                                       valid_data_iterator, model,
771
772
                                       iteration, process_non_loss_data_func,
                                       False)
773

774
775
        # Checkpointing
        saved_checkpoint = False
776
777
778
779
        if args.exit_signal_handler:
            signal_handler = get_signal_handler()
            if any(signal_handler.signals_received()):
                save_checkpoint_and_time(iteration, model, optimizer,
780
                                         opt_param_scheduler)
781
782
783
                print_datetime('exiting program after receiving SIGTERM.')
                sys.exit()

784
785
786
        if args.save and args.save_interval and \
           iteration % args.save_interval == 0:
            save_checkpoint_and_time(iteration, model, optimizer,
787
                                     opt_param_scheduler)
788
789
            saved_checkpoint = True

790
791
792
793
794
795
796
797
798
799
800
        # 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,
801
                                             opt_param_scheduler)
802
                print_datetime('exiting program after {} minutes'.format(train_time))
803
804
                sys.exit()

805
        # Exiting based on iterations
806
        if args.exit_interval and iteration % args.exit_interval == 0:
807
808
            if not saved_checkpoint:
                save_checkpoint_and_time(iteration, model, optimizer,
809
                                         opt_param_scheduler)
810
            torch.distributed.barrier()
811
            print_datetime('exiting program at iteration {}'.format(iteration))
Mohammad's avatar
Mohammad committed
812
            sys.exit()
813

814

mohammad's avatar
mohammad committed
815
    return iteration
816
817


818
819
820
821
822
def evaluate(forward_step_func,
             data_iterator,
             model,
             process_non_loss_data_func,
             verbose=False):
823
    """Evaluation."""
Mohammad's avatar
Mohammad committed
824
    args = get_args()
825

Vijay Korthikanti's avatar
Vijay Korthikanti committed
826
827
    if args.vision_pretraining and args.vision_pretraining_type == "dino":
        compute_feature_bank(model)
828

829
    # Turn on evaluation mode which disables dropout.
830
831
    for model_module in model:
        model_module.eval()
832
833
834
835
836
837
838
839
840
841

    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))
842

843
            forward_backward_func = get_forward_backward_func()
844
845
846
847
            loss_dicts = forward_backward_func(
                forward_step_func, data_iterator, model, optimizer=None,
                timers=None, forward_only=True)

848
            # Empty unused memory
Lawrence McAfee's avatar
Lawrence McAfee committed
849
            if args.empty_unused_memory_level >= 1:
850
851
                torch.cuda.empty_cache()

852
853
854
            if mpu.is_pipeline_last_stage(ignore_virtual=True):
                # Reduce across processes.
                for loss_dict in loss_dicts:
855
                    for key in loss_dict:
856
857
                        total_loss_dict[key] = total_loss_dict.get(
                            key, torch.cuda.FloatTensor([0.0])) + loss_dict[key]
858

859
            args.consumed_valid_samples += mpu.get_data_parallel_world_size() \
860
                                           * args.micro_batch_size \
mohammad's avatar
mohammad committed
861
                                           * get_num_microbatches()
862
863
864
865
866
867
        collected_non_loss_data = None
        if process_non_loss_data_func is not None and is_last_rank():
            collected_non_loss_data = forward_backward_func(
                forward_step_func, data_iterator, model, optimizer=None,
                timers=None, forward_only=True, collect_non_loss_data=True)

868
    # Move model back to the train mode.
869
870
    for model_module in model:
        model_module.train()
871
872

    for key in total_loss_dict:
mohammad's avatar
mohammad committed
873
        total_loss_dict[key] /= args.eval_iters * get_num_microbatches()
874

875
    return total_loss_dict, collected_non_loss_data
876
877
878

def evaluate_and_print_results(prefix, forward_step_func,
                               data_iterator, model,
879
880
                               iteration, process_non_loss_data_func,
                               verbose=False):
881
    """Helper function to evaluate and dump results on screen."""
882
    args = get_args()
Mohammad's avatar
Mohammad committed
883
884
    writer = get_tensorboard_writer()

885
886
887
    total_loss_dict, collected_non_loss_data = evaluate(
        forward_step_func, data_iterator, model,
        process_non_loss_data_func, verbose)
888
889
890
891
892
    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
893
        if writer:
mohammad's avatar
mohammad committed
894
            writer.add_scalar('{} validation'.format(key),
895
896
                              total_loss_dict[key].item(),
                              iteration)
mohammad's avatar
mohammad committed
897
            writer.add_scalar('{} validation vs samples'.format(key),
898
899
                              total_loss_dict[key].item(),
                              args.consumed_train_samples)
900
            if args.log_validation_ppl_to_tensorboard:
mohammad's avatar
mohammad committed
901
                writer.add_scalar('{} validation ppl'.format(key), ppl,
902
                                  iteration)
mohammad's avatar
mohammad committed
903
                writer.add_scalar('{} validation ppl vs samples'.format(key),
904
                                  ppl, args.consumed_train_samples)
905

906
907
908
    if process_non_loss_data_func is not None and writer and is_last_rank():
        process_non_loss_data_func(collected_non_loss_data, iteration, writer)

909
    length = len(string) + 1
910
911
912
    print_rank_last('-' * length)
    print_rank_last(string)
    print_rank_last('-' * length)
913
914


Vijay Korthikanti's avatar
Vijay Korthikanti committed
915
def cyclic_iter(iter):
916
    while True:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
917
        for x in iter:
918
919
            yield x

920
921
922
def build_train_valid_test_data_iterators(
        build_train_valid_test_datasets_provider):
    """XXX"""
Mohammad's avatar
Mohammad committed
923
    args = get_args()
924

925
926
927
    (train_dataloader, valid_dataloader, test_dataloader) = (None, None, None)

    print_rank_0('> building train, validation, and test datasets ...')
928
929
930

    # Backward compatibility, assume fixed batch size.
    if args.iteration > 0 and args.consumed_train_samples == 0:
931
932
        assert args.train_samples is None, \
            'only backward compatiblity support for iteration-based training'
mohammad's avatar
mohammad committed
933
        args.consumed_train_samples = args.iteration * args.global_batch_size
934
    if args.iteration > 0 and args.consumed_valid_samples == 0:
935
936
937
        if args.train_samples is None:
            args.consumed_valid_samples = (args.iteration // args.eval_interval) * \
                args.eval_iters * args.global_batch_size
938

939
    # Data loader only on rank 0 of each model parallel group.
940
    if mpu.get_tensor_model_parallel_rank() == 0:
941
942

        # Number of train/valid/test samples.
943
944
945
946
947
948
        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
949
        test_iters = args.eval_iters
950
        train_val_test_num_samples = [train_samples,
mohammad's avatar
mohammad committed
951
952
                                      eval_iters * args.global_batch_size,
                                      test_iters * args.global_batch_size]
953
954
955
956
957
958
959
960
961
962
        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.
963
964
965
966
967
        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)
968
969
970
971
972
973
974
975
976
977
978
979
980

        # 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,
981
982
                                mpu.get_tensor_model_parallel_src_rank(),
                                group=mpu.get_tensor_model_parallel_group())
983
984
985
986
987
    args.do_train = flags[0].item()
    args.do_valid = flags[1].item()
    args.do_test = flags[2].item()

    # Build iterators.
Vijay Korthikanti's avatar
Vijay Korthikanti committed
988
989
990
    dl_type = args.dataloader_type
    assert dl_type in ['single', 'cyclic']

991
    if train_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
992
993
        train_data_iterator = iter(train_dataloader) if dl_type == 'single' \
                              else iter(cyclic_iter(train_dataloader))
994
995
996
    else:
        train_data_iterator = None

997
    if valid_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
998
999
        valid_data_iterator = iter(valid_dataloader) if dl_type == 'single' \
                              else iter(cyclic_iter(valid_dataloader))
1000
    else:
1001
        valid_data_iterator = None
1002

1003
    if test_dataloader is not None:
Vijay Korthikanti's avatar
Vijay Korthikanti committed
1004
1005
        test_data_iterator = iter(test_dataloader) if dl_type == 'single' \
                             else iter(cyclic_iter(test_dataloader))
1006
1007
1008
    else:
        test_data_iterator = None

1009
    return train_data_iterator, valid_data_iterator, test_data_iterator