optimizer.py 28.9 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# coding=utf-8
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     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.

"""Megatron optimizer."""
mohammad's avatar
mohammad committed
17
18
19
20
21

from abc import ABC
from abc import abstractmethod
from apex.multi_tensor_apply import multi_tensor_applier
import amp_C
22
23
import torch
from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
mohammad's avatar
mohammad committed
24

mohammad's avatar
mohammad committed
25
26
from megatron import get_timers
from megatron import mpu
mohammad's avatar
mohammad committed
27
from megatron import print_rank_0
28
29
30
from megatron.model import DistributedDataParallel as LocalDDP
from megatron.model import Float16Module
from megatron.utils import unwrap_model
mohammad's avatar
mohammad committed
31

Rewon Child's avatar
Rewon Child committed
32
from .clip_grads import clip_grad_norm_fp32, count_zeros_fp32
mohammad's avatar
mohammad committed
33

34
35
# >>>
from lutil import pax, tp
36

Lawrence McAfee's avatar
Lawrence McAfee committed
37
DEBUG_ITERATION = 1 # 10
38
# <<<
mohammad's avatar
mohammad committed
39

Lawrence McAfee's avatar
Lawrence McAfee committed
40

mohammad's avatar
mohammad committed
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
def _zero_grad_group_helper(group, set_to_none):
    """Zero out the gradient for a group of parameters.
    Note: copied from torch.optim.optimizer."""
    for param in group:
        if param.grad is not None:
            if set_to_none:
                param.grad = None
            else:
                if param.grad.grad_fn is not None:
                    param.grad.detach_()
                else:
                    param.grad.requires_grad_(False)
                param.grad.zero_()


56
def _multi_tensor_copy_this_to_that(this, that, overflow_buf=None):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
57
58
59
60
    """Use multi-tensor-applier to copy values from one list to another.
    We don't have a blfoat16 implementation so for now if the overflow_buf
    is not provided, we default back to simple loop copy to be compatible
    with bfloat16."""
61
62
    if overflow_buf:
        overflow_buf.fill_(0)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
63
64
65
66
67
        # Scaling with factor `1.0` is equivalent to copy.
        multi_tensor_applier(amp_C.multi_tensor_scale,
                             overflow_buf,
                             [this, that],
                             1.0)
68
    else:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
69
70
71
        for this_, that_ in zip(this, that):
            that_.copy_(this_)

72

mohammad's avatar
mohammad committed
73
74
75

class MegatronOptimizer(ABC):

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
76
77
78

    def __init__(self, optimizer, clip_grad,
                 log_num_zeros_in_grad,
79
                 params_have_main_grad,
80
81
                 use_contiguous_buffers_in_local_ddp,
                 models):
82

mohammad's avatar
mohammad committed
83
84
85
        """Input optimizer is the base optimizer for example Adam."""
        self.optimizer = optimizer
        assert self.optimizer, 'no optimizer is provided.'
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
86
87
88
89
        # Set gradient clipping and logging params.
        self.clip_grad = clip_grad
        self.log_num_zeros_in_grad = log_num_zeros_in_grad
        self.params_have_main_grad = params_have_main_grad
90
        self.use_contiguous_buffers_in_local_ddp = use_contiguous_buffers_in_local_ddp
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
91

92
93
94
95
        # 'models' are retained for access to the contiguous grad buffers.
        # (see distributed optimizer)
        self.models = models

96
        if self.use_contiguous_buffers_in_local_ddp:
97
98
            assert self.params_have_main_grad, \
                "use of contiguous buffer requires that params have main grad"
mohammad's avatar
mohammad committed
99

Rewon Child's avatar
Rewon Child committed
100
    def get_parameters(self):
101
102
103
104
        params = []
        for param_group in self.optimizer.param_groups:
            for param in param_group['params']:
                params.append(param)
Rewon Child's avatar
Rewon Child committed
105
106
        return params

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
107

108
109
110
111
112
    def get_model_parallel_group(self):
        '''Default returned here, but the distributed optimizer overrides this.'''
        return mpu.get_model_parallel_group()


Lawrence McAfee's avatar
Lawrence McAfee committed
113
114
    def clip_grad_norm(self, clip_grad, ITERATION):
        params = self.get_parameters()
115
116
117
118
        return clip_grad_norm_fp32(
            params, clip_grad,
            model_parallel_group=self.get_model_parallel_group(),
            ITERATION = ITERATION)
119

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
120

Rewon Child's avatar
Rewon Child committed
121
122
    def count_zeros(self):
        params = self.get_parameters()
123
124
        return count_zeros_fp32(params,
                                model_parallel_group=self.get_model_parallel_group())
Rewon Child's avatar
Rewon Child committed
125

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
126

mohammad's avatar
mohammad committed
127
128
129
130
    @abstractmethod
    def zero_grad(self, set_to_none=True):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
131

mohammad's avatar
mohammad committed
132
133
    @abstractmethod
    def get_loss_scale(self):
134
        """The output should be a cuda tensor of size 1."""
mohammad's avatar
mohammad committed
135
136
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
137

mohammad's avatar
mohammad committed
138
139
140
141
    def scale_loss(self, loss):
        """Simple scaling."""
        return self.get_loss_scale() * loss

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
142

143
144
    @abstractmethod
    def reload_model_params(self):
145
146
147
148
149
        """Refreshes any internal state from the current model parameters.
        Call whenever the parameters are changed outside of the optimizer.
        For example, when we load a model from a checkpoint  without loading
        the optimizer, the model parameters are updated but for fp16 optimizer
        with main parameters, the main parameters need to also be updated."""
150
151
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
152

mohammad's avatar
mohammad committed
153
154
155
156
    @abstractmethod
    def state_dict(self):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
157

mohammad's avatar
mohammad committed
158
159
160
161
    @abstractmethod
    def load_state_dict(self, state_dict):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
162

mohammad's avatar
mohammad committed
163
164
165
166
167
168
169
170
171
172
    # Promote state so it can be retrieved or set via
    # "optimizer_instance.state"
    def _get_state(self):
        return self.optimizer.state

    def _set_state(self, value):
        self.optimizer.state = value

    state = property(_get_state, _set_state)

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
173

mohammad's avatar
mohammad committed
174
175
176
177
178
179
180
181
182
183
184
185
    # Promote param_groups so it can be retrieved or set via
    # "optimizer_instance.param_groups"
    # (for example, to adjust the learning rate)
    def _get_param_groups(self):
        return self.optimizer.param_groups

    def _set_param_groups(self, value):
        self.optimizer.param_groups = value

    param_groups = property(_get_param_groups, _set_param_groups)


186
    @abstractmethod
187
    def step(self, args, timers):
188
189
        pass

190
191
192
    def gather_model_params(self, args, timers, ITERATION):
        '''For the case of a non-distributed-optimizer, there is nothing to
        do here.'''
193
194
        pass

195
    def allreduce_word_embedding_grads(self, args):
196
197
        '''
        All-reduce word embedding grads.
198

199
200
201
202
        Reduce grads 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).
        '''
203
204
205
206

        if mpu.is_rank_in_embedding_group(ignore_virtual=True) and \
                mpu.get_pipeline_model_parallel_world_size() > 1:
            if mpu.is_pipeline_first_stage(ignore_virtual=True):
207
                unwrapped_model = self.models[0]
208
            elif mpu.is_pipeline_last_stage(ignore_virtual=True):
209
                unwrapped_model = self.models[-1]
210
            else:  # We do not support the interleaved schedule for T5 yet.
211
                unwrapped_model = self.models[0]
212
213
214
215
216
217
218
219
220
221
222
            unwrapped_model = unwrap_model(
                unwrapped_model, (torchDDP, LocalDDP, Float16Module))

            if unwrapped_model.share_word_embeddings:
                word_embeddings_weight = unwrapped_model.word_embeddings_weight()
                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())

223
    def allreduce_position_embedding_grads(self, args):
224
225
226
227
228
229
        '''
        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.
        '''
230
231
232
        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:
233
            unwrapped_model = self.models[0]
234
235
236
237
238
239
            unwrapped_model = unwrap_model(
                unwrapped_model, (torchDDP, LocalDDP, Float16Module))
            assert args.DDP_impl == 'local', \
                'T5 model is only supported with local DDP mode'
            grad = unwrapped_model.language_model.embedding.position_embeddings.weight.main_grad
            torch.distributed.all_reduce(grad, group=mpu.get_position_embedding_group())
240

241
    def allreduce_embedding_grads(self, args):
Lawrence McAfee's avatar
Lawrence McAfee committed
242
243
244
        # >>>
        # return # ** .. TEMPORARY .. **
        # <<<
245
246
        self.allreduce_word_embedding_grads(args)
        self.allreduce_position_embedding_grads(args)
247

248
    def reduce_model_grads(self, args, timers):
249
250
251
252

        # All-reduce if needed.
        if args.DDP_impl == 'local':
            timers('backward-params-all-reduce').start()
253
254
            for model in self.models:
                model.allreduce_gradients()
255
256
257
258
            timers('backward-params-all-reduce').stop()

        # All-reduce embedding grads.
        timers('backward-embedding-all-reduce').start()
259
        self.allreduce_embedding_grads(args)
260
261
        timers('backward-embedding-all-reduce').stop()

262

263
264
# class BaseFloat16Optimizer(MegatronOptimizer):
class MixedPrecisionOptimizer(MegatronOptimizer):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
265
266

    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
267
                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
268
269
                 bf16, grad_scaler,
                 models):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
270

Lawrence McAfee's avatar
Lawrence McAfee committed
271
        super().__init__(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
272
            optimizer, clip_grad, log_num_zeros_in_grad,
273
274
            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
            models)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
275
276

        self.bf16 = bf16
mohammad's avatar
mohammad committed
277
        self.grad_scaler = grad_scaler
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
278
279
280
        # None grad scaler is only supported for bf16.
        if self.grad_scaler is None:
            assert self.bf16, 'fp16 expects a grad scaler.'
mohammad's avatar
mohammad committed
281
282
283

        # Tensor used to determine if a nan/if has happend.
        # Any non-zero value indicates inf/nan.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
284
285
286
287
        # Note that we keep this for the cases that grad scaler is none.
        # We still record nan/inf if we have a bfloat16 with a grad scaler.
        if self.grad_scaler:
            self.found_inf = torch.cuda.FloatTensor([0.0])
mohammad's avatar
mohammad committed
288
289

        # Dummy tensor needed for apex multi-apply tensor.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
290
291
292
293
294
295
296
297
298
299
        # For bfloat, we don't have multi-tensor apply and for now
        # we set it to none so the multi-tensor apply gets ignored.
        if bf16:
            self._dummy_overflow_buf = None
        else:
            self._dummy_overflow_buf = torch.cuda.IntTensor([0])

        # In case grad scaler is not passed, define the unity scale.
        if self.grad_scaler is None:
            self._scale_one = torch.cuda.FloatTensor([1.0])
mohammad's avatar
mohammad committed
300

Lawrence McAfee's avatar
Lawrence McAfee committed
301
302
303
304
305
306
307

    def get_loss_scale(self):
        if self.grad_scaler is None:
            return self._scale_one
        return self.grad_scaler.scale


Lawrence McAfee's avatar
Lawrence McAfee committed
308
309
310
311
    def reload_model_params(self):
        self._copy_model_params_to_main_params()


312
    def _unscale_main_grads_and_check_for_nan(self):
Lawrence McAfee's avatar
Lawrence McAfee committed
313
314
315
316
317
318
319
320
321
322
323
324
325
326

        # Collect main grads.
        main_grads = self._collect_main_grad_data_for_unscaling()
        # pax(1, {"main_grads": main_grads})

        # Reset found inf.
        self.found_inf.fill_(0.0)

        # Unscale and set found inf/nan
        torch._amp_foreach_non_finite_check_and_unscale_(
            main_grads, self.found_inf, self.grad_scaler.inv_scale)

        # Update across all model parallel instances.
        torch.distributed.all_reduce(self.found_inf,
327
328
                                     op=torch.distributed.ReduceOp.MAX,
                                     group=self.get_model_parallel_group())
Lawrence McAfee's avatar
Lawrence McAfee committed
329
330
331
332

        # Check for nan.
        found_inf_flag = (self.found_inf.item() > 0)

333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
        # >>>
        # if self.grad_scaler.scale <= 131072:
        #     pax(0, {
        #         # "grad_scaler" : self.grad_scaler,
        #         # "found_inf_flag" : found_inf_flag,
        #         "model_params" : [
        #             p
        #             for m in self.models
        #             for p in m.parameters()
        #         ],
        #         "model_grads" : [
        #             p.main_grad
        #             for m in self.models
        #             for p in m.parameters()
        #         ],
        #         # "main_grads" : main_grads,
        #     })
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
352
353
        return found_inf_flag

Lawrence McAfee's avatar
Lawrence McAfee committed
354
    # >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
Lawrence McAfee's avatar
Lawrence McAfee committed
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
    @classmethod
    def debug_base(cls, ITERATION, key, value):
        from megatron import get_args
        args = get_args()
        my_rank = torch.distributed.get_rank()
        if ITERATION != DEBUG_ITERATION:
            return
        for r in range(torch.distributed.get_world_size()):
            if my_rank == r:
                # prefix = "            + "
                prefix = ""
                print("%sbr/%s; [r%d, i%d]; %s, %.12e" % (prefix, "fix " if args.use_distributed_optimizer else "main", my_rank, ITERATION, key, value))
            torch.distributed.barrier()
        torch.distributed.barrier()
        # if my_rank == 0:
        #     raise Exception("debug.")
        # else:
        #     exit(0)
        exit(0)
    def debug_model(self, ITERATION, key, use_grad):
        use_grad = bool(use_grad)
        tensors = [
            (p.main_grad.float() if use_grad else p.float())
            for m in self.models for p in m.parameters()
        ]
        count = sum(t.nelement() for t in tensors)
        return self.debug_base(
            ITERATION,
            "model/%s, %s [count %d]" % (
                "grad" if use_grad else "param",
                key,
                count,
            ),
            # sum(torch.sum(torch.abs(t)) for t in tensors).item() / count,
            sum(torch.sum(torch.abs(t)) for t in tensors),
        )
    def debug_main(self, ITERATION, key, use_grad):
        use_grad = bool(use_grad)
        tensors = [
            p.grad if use_grad else p
            for g in self.optimizer.param_groups
            for p in g["params"]
        ]
        tensors = [ t.float() for t in tensors ]
        count = sum(t.nelement() for t in tensors)
        return self.debug_base(
            ITERATION,
            "main/%s, %s [count %d]" % (
                "grad" if use_grad else "param",
                key,
                count,
            ),
            sum(torch.sum(torch.abs(t)) for t in tensors),
        )
Lawrence McAfee's avatar
Lawrence McAfee committed
409
    # <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
Lawrence McAfee's avatar
Lawrence McAfee committed
410
411

    @torch.no_grad()
412
    def step(self, args, timers, ITERATION):
Lawrence McAfee's avatar
Lawrence McAfee committed
413

414
        # >>>
415
416
        # self.debug_model(ITERATION, "before copy grad.", 0)
        # self.debug_main(ITERATION, "before copy grad.", 0)
417
418
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
419
420
        # Copy gradients from model params to main params.
        timers('optimizer-copy-to-main-grad').start()
421
        self._copy_model_grads_to_main_grads(ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
422
423
424
425
426
427
428
429
430
431
432
        timers('optimizer-copy-to-main-grad').stop()

        # Do unscale, check for inf, and update grad scaler only for
        # the case that grad scaler is provided.
        if self.grad_scaler:

            # Unscale and check for inf/nan.
            timers('optimizer-unscale-and-check-inf').start()
            found_inf_flag = self._unscale_main_grads_and_check_for_nan()
            timers('optimizer-unscale-and-check-inf').stop()

433
434
435
436
            # >>>
            
            # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
437
438
439
440
441
442
443
444
445
446
447
448
            # We are done with scaling gradients
            # so we can update the loss scale.
            self.grad_scaler.update(found_inf_flag)

            # If we found inf/nan, skip the update.
            if found_inf_flag:
                return False, None, None

        # Clip the main gradients.
        timers('optimizer-clip-main-grad').start()
        grad_norm = None
        if self.clip_grad > 0.0:
Lawrence McAfee's avatar
Lawrence McAfee committed
449
            grad_norm = self.clip_grad_norm(self.clip_grad, ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
450
451
452
453
454
455
        timers('optimizer-clip-main-grad').stop()

        # count the zeros in the grads
        num_zeros_in_grad = self.count_zeros() if \
                            self.log_num_zeros_in_grad else None

456
457
458
        # Step the optimizer.
        self.optimizer.step()

Lawrence McAfee's avatar
Lawrence McAfee committed
459
460
        # Update params from main params.
        timers('optimizer-copy-main-to-model-params').start()
461
        self._copy_main_params_to_model_params(ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
462
463
        timers('optimizer-copy-main-to-model-params').stop()

Lawrence McAfee's avatar
Lawrence McAfee committed
464
465
466
467
468
        # >>>
        # self.debug_model(ITERATION, "after copy param.", 0)
        # self.debug_main(ITERATION, "after copy param.", 0)
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
469
470
471
472
        # Successful update.
        return True, grad_norm, num_zeros_in_grad


Lawrence McAfee's avatar
Lawrence McAfee committed
473
# class Float16OptimizerWithFloat16Params(MegatronOptimizer):
474
475
# class Float16OptimizerWithFloat16Params(BaseFloat16Optimizer):
class Float16OptimizerWithFloat16Params(MixedPrecisionOptimizer):
Lawrence McAfee's avatar
Lawrence McAfee committed
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
    """Float16 optimizer for fp16 and bf16 data types.

    Arguments:
        optimizer: base optimizer such as Adam or SGD
        clip_grad: clip gradeints with this global L2 norm. Note
            that clipping is ignored if clip_grad == 0
        log_num_zeros_in_grad: return number of zeros in the gradients.
        params_have_main_grad: flag indicating if parameters have
            a `main_grad` field. If this is set, we are assuming
            that the model parameters are store in the `main_grad`
            field instead of the typical `grad` field. This happens
            for the DDP cases where there is a continuous buffer
            holding the gradients. For example for bfloat16, we want
            to do gradient accumulation and all-reduces in float32
            and as a result we store those gradients in the main_grad.
            Note that main grad is not necessarily in float32.
        bf16: if true, the model is running in bfloat16.
        grad_scaler: used for scaling gradients. Note that this can be
            None. This case happens when `bf16 = True` and we don't
            use any loss scale. Note that for `bf16 = True`, we can have
            a constnat gradient scaler. Also for `bf16 = False`, we
            always require a grad scaler.
    """

    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
Lawrence McAfee's avatar
Lawrence McAfee committed
502
                 bf16, grad_scaler, models):
Lawrence McAfee's avatar
Lawrence McAfee committed
503
504
505
506

        super().__init__(
            optimizer, clip_grad, log_num_zeros_in_grad,
            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
Lawrence McAfee's avatar
Lawrence McAfee committed
507
            bf16, grad_scaler, models)
Lawrence McAfee's avatar
Lawrence McAfee committed
508

mohammad's avatar
mohammad committed
509
        # ======================
510
        # main parameter stuff
mohammad's avatar
mohammad committed
511
512
513
        # ======================

        # Three groups of parameters:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
514
515
        #   float16_groups: original float16 parameters
        #   fp32_from_float16_groups: fp32 copy of float16 parameters
mohammad's avatar
mohammad committed
516
        #   fp32_from_fp32_groups: original fp32 parameters
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
517
518
        self.float16_groups = []
        self.fp32_from_float16_groups = []
mohammad's avatar
mohammad committed
519
520
521
522
        self.fp32_from_fp32_groups = []

        # For all the groups in the original optimizer:
        for param_group in self.optimizer.param_groups:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
523
            float16_params_this_group = []
mohammad's avatar
mohammad committed
524
            fp32_params_this_group = []
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
525
            fp32_from_float16_params_this_group = []
mohammad's avatar
mohammad committed
526
527
528
529
            # For all the parameters in this group:
            for i, param in enumerate(param_group['params']):
                if param.requires_grad:

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
530
531
532
533
                    # float16 params:
                    if param.type() in ['torch.cuda.HalfTensor',
                                        'torch.cuda.BFloat16Tensor']:
                        float16_params_this_group.append(param)
mohammad's avatar
mohammad committed
534
                        # Create a copy
535
                        main_param = param.detach().clone().float()
mohammad's avatar
mohammad committed
536
                        # Copy tensor model parallel attributes.
537
                        mpu.copy_tensor_model_parallel_attributes(main_param,
mohammad's avatar
mohammad committed
538
                                                                  param)
539
                        if hasattr(param, 'shared'):
540
                            main_param.shared = param.shared
mohammad's avatar
mohammad committed
541
                        # Replace the optimizer params with the new fp32 copy.
542
                        param_group['params'][i] = main_param
543

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
544
                        fp32_from_float16_params_this_group.append(main_param)
545
                        # Reset existing state dict key to the new main param.
mohammad's avatar
mohammad committed
546
                        if param in self.optimizer.state:
547
                            self.optimizer.state[main_param] \
mohammad's avatar
mohammad committed
548
549
550
551
552
553
554
555
                                = self.optimizer.state.pop(param)

                    # fp32 params.
                    elif param.type() == 'torch.cuda.FloatTensor':
                        fp32_params_this_group.append(param)
                        param_group['params'][i] = param

                    else:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
556
557
558
559
560
561
562
563
564
                        raise TypeError('Wrapped parameters must be one of '
                                        'torch.cuda.FloatTensor,  '
                                        'torch.cuda.HalfTensor, or '
                                        'torch.cuda.BFloat16Tensor. '
                                        'Received {}'.format(param.type()))

            self.float16_groups.append(float16_params_this_group)
            self.fp32_from_float16_groups.append(
                fp32_from_float16_params_this_group)
mohammad's avatar
mohammad committed
565
566
567
568
569
570
571
572
573
            self.fp32_from_fp32_groups.append(fp32_params_this_group)

        # Leverage state_dict() and load_state_dict() to
        # recast preexisting per-param state tensors
        self.optimizer.load_state_dict(self.optimizer.state_dict())


    def zero_grad(self, set_to_none=True):
        """We only need to zero the model related parameters, i.e.,
574
575
576
577
        float16_groups & fp32_from_fp32_groups. We additionally zero
        fp32_from_float16_groups as a memory optimization to reduce
        fragmentation; in the case of set_to_none==True, the space
        used by this field can be safely deallocated at this point."""
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
578
        for group in self.float16_groups:
mohammad's avatar
mohammad committed
579
            _zero_grad_group_helper(group, set_to_none)
580
581
        for group in self.fp32_from_float16_groups:
            _zero_grad_group_helper(group, set_to_none)
mohammad's avatar
mohammad committed
582
583
584
585
        for group in self.fp32_from_fp32_groups:
            _zero_grad_group_helper(group, set_to_none)


586
    def _collect_main_grad_data_for_unscaling(self):
587

588
        main_grads = []
589

590
591
592
593
594
        # fp32 params from float16 ones.
        for main_group in self.fp32_from_float16_groups:
            for main_param in main_group:
                if main_param.grad is not None:
                    main_grads.append(main_param.grad.data)
595

596
597
598
599
600
601
602
        # Append fp32 parameters.
        for main_group in self.fp32_from_fp32_groups:
            for main_param in main_group:
                if main_param.grad is not None:
                    main_grads.append(main_param.grad.data)
        
        return main_grads
603
604


605
606
607
608
609
610
611
612
613
    def _get_model_and_main_params_data_float16(self):
        model_data = []
        main_data = []
        for model_group, main_group in zip(self.float16_groups,
                                           self.fp32_from_float16_groups):
            for model_param, main_param in zip(model_group, main_group):
                model_data.append(model_param.data)
                main_data.append(main_param.data)
        return model_data, main_data
614

Lawrence McAfee's avatar
Lawrence McAfee committed
615

616
    def _copy_model_grads_to_main_grads(self, ITERATION):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
617
618
619
        # This only needs to be done for the float16 group.
        for model_group, main_group in zip(self.float16_groups,
                                           self.fp32_from_float16_groups):
620
            for model_param, main_param in zip(model_group, main_group):
621
                if self.params_have_main_grad and hasattr(model_param, 'main_grad'):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
622
623
624
625
                    main_param.grad = model_param.main_grad.float()
                else:
                    if model_param.grad is not None:
                        main_param.grad = model_param.grad.float()
626
627
628
629
630

                # Safe to deallocate model's grad/main_grad after copying.
                # (If using contiguous buffers, main_grad's memory should
                # persist and therefore should not be deallocated.)
                model_param.grad = None
631
                if self.params_have_main_grad and \
632
                   not self.use_contiguous_buffers_in_local_ddp:
633
634
                    model_param.main_grad = None

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
635
636
637
638
639
        # For fp32 grads, we need to reset the grads to main grad.
        if self.params_have_main_grad:
            for model_group in self.fp32_from_fp32_groups:
                for model_param in model_group:
                    model_param.grad = model_param.main_grad
mohammad's avatar
mohammad committed
640

641
642
643
                    # Safe to de-reference model's main_grad after copying.
                    # (If using contiguous buffers, main_grad's memory should
                    # persist and therefore should not be deallocated.)
644
                    if not self.use_contiguous_buffers_in_local_ddp:
645
                        model_param.main_grad = None
mohammad's avatar
mohammad committed
646

647

648
    def _copy_main_params_to_model_params(self, ITERATION):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
649
650
        # Only needed for the float16 params.
        model_data, main_data = self._get_model_and_main_params_data_float16()
651
652
653
654
655
        _multi_tensor_copy_this_to_that(this=main_data, that=model_data,
                                        overflow_buf=self._dummy_overflow_buf)


    def _copy_model_params_to_main_params(self):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
656
657
        # Only needed for the float16 params.
        model_data, main_data = self._get_model_and_main_params_data_float16()
658
659
        _multi_tensor_copy_this_to_that(this=model_data, that=main_data,
                                        overflow_buf=self._dummy_overflow_buf)
660
661


mohammad's avatar
mohammad committed
662
663
664
    def state_dict(self):
        state_dict = {}
        state_dict['optimizer'] = self.optimizer.state_dict()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
665
666
667
        if self.grad_scaler:
            state_dict['grad_scaler'] = self.grad_scaler.state_dict()
        state_dict['fp32_from_fp16_params'] = self.fp32_from_float16_groups
mohammad's avatar
mohammad committed
668
669
670
671
        return state_dict


    def load_state_dict(self, state_dict):
mohammad's avatar
mohammad committed
672
673
674
675
676
677
678
679
680
681
682
683
684
        # Optimizer.
        optimizer_key = 'optimizer'
        if optimizer_key not in state_dict:
            optimizer_key = 'optimizer_state_dict'
            print_rank_0('***WARNING*** loading optimizer from '
                         'an old checkpoint ...')
        self.optimizer.load_state_dict(state_dict[optimizer_key])

        # Grad scaler.
        if 'grad_scaler' not in state_dict:
            print_rank_0('***WARNING*** found an old checkpoint, will not '
                         'load grad scaler ...')
        else:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
685
686
687
688
689
690
            if self.grad_scaler:
                self.grad_scaler.load_state_dict(state_dict['grad_scaler'])
            else:
                print_rank_0('***WARNING*** fould the grad scaler in the '
                             'checkpoint but it is None in the class. '
                             'Skipping loading grad scaler ...')
mohammad's avatar
mohammad committed
691

692
        # Copy data for the main params.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
693
694
695
        fp32_from_float16_params_key = 'fp32_from_fp16_params'
        if fp32_from_float16_params_key not in state_dict:
            fp32_from_float16_params_key = 'fp32_from_fp16'
mohammad's avatar
mohammad committed
696
        for current_group, saved_group in zip(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
697
698
                self.fp32_from_float16_groups,
                state_dict[fp32_from_float16_params_key]):
mohammad's avatar
mohammad committed
699
700
701
702
            for current_param, saved_param in zip(current_group, saved_group):
                current_param.data.copy_(saved_param.data)


mohammad's avatar
mohammad committed
703
704
class FP32Optimizer(MegatronOptimizer):

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
705
706
    def __init__(self, optimizer, clip_grad,
                 log_num_zeros_in_grad,
707
                 params_have_main_grad,
708
709
                 use_contiguous_buffers_in_local_ddp,
                 models):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
710
711
712

        super(FP32Optimizer, self).__init__(
            optimizer, clip_grad, log_num_zeros_in_grad,
713
714
            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
            models)
mohammad's avatar
mohammad committed
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730

        self._scale = torch.cuda.FloatTensor([1.0])


    def zero_grad(self, set_to_none=True):
        """Copied from torch.optim.optimizer"""
        for group in self.optimizer.param_groups:
            _zero_grad_group_helper(group['params'], set_to_none)


    def get_loss_scale(self):
        """FP32 optimizer does not do any scaling."""
        return self._scale


    @torch.no_grad()
731
    def step(self, args, timers, ITERATION):
mohammad's avatar
mohammad committed
732
        """Clip gradients (if needed) and step the base optimizer.
mohammad's avatar
mohammad committed
733
        Always return successful since there is no overflow."""
mohammad's avatar
mohammad committed
734

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
735
736
737
738
739
740
        # Copy main_grads to grads.
        if self.params_have_main_grad:
            for param_group in self.optimizer.param_groups:
                for param in param_group['params']:
                    param.grad = param.main_grad

741
742
743
                    # Safe to de-reference model's main_grad after copying.
                    # (If using contiguous buffers, main_grad's memory should
                    # persist and therefore should not be deallocated.)
744
                    if not self.use_contiguous_buffers_in_local_ddp:
745
746
                        param.main_grad = None

mohammad's avatar
mohammad committed
747
        # Clip gradients.
748
        grad_norm = None
mohammad's avatar
mohammad committed
749
        if self.clip_grad > 0.0:
750
            grad_norm = self.clip_grad_norm(self.clip_grad, ITERATION)
mohammad's avatar
mohammad committed
751

Rewon Child's avatar
Rewon Child committed
752
        # count the zeros in the grads
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
753
754
        num_zeros_in_grad = self.count_zeros() if \
                            self.log_num_zeros_in_grad else None
Rewon Child's avatar
Rewon Child committed
755

mohammad's avatar
mohammad committed
756
757
758
759
        # Update parameters.
        self.optimizer.step()

        # No overflow for FP32 optimizer.
760
        return True, grad_norm, num_zeros_in_grad
mohammad's avatar
mohammad committed
761
762


763
764
765
766
    def reload_model_params(self):
        pass


mohammad's avatar
mohammad committed
767
768
769
770
771
772
    def state_dict(self):
        return self.optimizer.state_dict()


    def load_state_dict(self, state_dict):
        self.optimizer.load_state_dict(state_dict)