"...experiment/trialdetail/ChangeColumnComponent.tsx" did not exist on "754b0043cf4b45d63151b8c5621650b00e12f7d9"
optimizer.py 29.7 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
22
23
24
25

from abc import ABC
from abc import abstractmethod

import torch

from apex.multi_tensor_apply import multi_tensor_applier
import amp_C

mohammad's avatar
mohammad committed
26
27
from megatron import get_timers
from megatron import mpu
mohammad's avatar
mohammad committed
28
29
from megatron import print_rank_0

Rewon Child's avatar
Rewon Child committed
30
from .clip_grads import clip_grad_norm_fp32, count_zeros_fp32
mohammad's avatar
mohammad committed
31

32
33
# >>>
from lutil import pax, tp
34

35
DEBUG_ITERATION = 2 # 10
36
# <<<
mohammad's avatar
mohammad committed
37

Lawrence McAfee's avatar
Lawrence McAfee committed
38

mohammad's avatar
mohammad committed
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
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_()


54
def _multi_tensor_copy_this_to_that(this, that, overflow_buf=None):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
55
56
57
58
    """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."""
59
60
    if overflow_buf:
        overflow_buf.fill_(0)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
61
62
63
64
65
        # Scaling with factor `1.0` is equivalent to copy.
        multi_tensor_applier(amp_C.multi_tensor_scale,
                             overflow_buf,
                             [this, that],
                             1.0)
66
    else:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
67
68
69
        for this_, that_ in zip(this, that):
            that_.copy_(this_)

70

mohammad's avatar
mohammad committed
71
72
73

class MegatronOptimizer(ABC):

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
74
75
76

    def __init__(self, optimizer, clip_grad,
                 log_num_zeros_in_grad,
77
                 params_have_main_grad,
78
                 use_contiguous_buffers_in_local_ddp):
79

mohammad's avatar
mohammad committed
80
81
82
        """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
83
84
85
86
        # 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
87
        self.use_contiguous_buffers_in_local_ddp = use_contiguous_buffers_in_local_ddp
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
88

89
        if self.use_contiguous_buffers_in_local_ddp:
90
91
            assert self.params_have_main_grad, \
                "use of contiguous buffer requires that params have main grad"
mohammad's avatar
mohammad committed
92

Rewon Child's avatar
Rewon Child committed
93
    def get_parameters(self):
94
95
96
97
        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
98
99
        return params

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
100

Lawrence McAfee's avatar
Lawrence McAfee committed
101
102
103
    def clip_grad_norm(self, clip_grad, ITERATION):
        params = self.get_parameters()
        return clip_grad_norm_fp32(params, clip_grad, ITERATION = ITERATION)
104

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
105

Rewon Child's avatar
Rewon Child committed
106
107
108
109
    def count_zeros(self):
        params = self.get_parameters()
        return count_zeros_fp32(params)

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
110

mohammad's avatar
mohammad committed
111
112
113
114
    @abstractmethod
    def zero_grad(self, set_to_none=True):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
115

mohammad's avatar
mohammad committed
116
117
    @abstractmethod
    def get_loss_scale(self):
118
        """The output should be a cuda tensor of size 1."""
mohammad's avatar
mohammad committed
119
120
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
121

mohammad's avatar
mohammad committed
122
123
124
125
    def scale_loss(self, loss):
        """Simple scaling."""
        return self.get_loss_scale() * loss

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
126

Lawrence McAfee's avatar
Lawrence McAfee committed
127
    @abstractmethod
128
    def reduce_grads(self):
Lawrence McAfee's avatar
Lawrence McAfee committed
129
130
131
        pass


mohammad's avatar
mohammad committed
132
133
134
135
    @abstractmethod
    def step(self):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
136

Lawrence McAfee's avatar
Lawrence McAfee committed
137
138
139
140
141
    @abstractmethod
    def gather_params(self):
        pass


142
143
    @abstractmethod
    def reload_model_params(self):
144
145
146
147
148
        """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."""
149
150
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
151

mohammad's avatar
mohammad committed
152
153
154
155
    @abstractmethod
    def state_dict(self):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
156

mohammad's avatar
mohammad committed
157
158
159
160
    @abstractmethod
    def load_state_dict(self, state_dict):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
161

mohammad's avatar
mohammad committed
162
163
164
165
166
167
168
169
170
171
    # 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
172

mohammad's avatar
mohammad committed
173
174
175
176
177
178
179
180
181
182
183
184
    # 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)


Lawrence McAfee's avatar
Lawrence McAfee committed
185
class BaseFloat16Optimizer(MegatronOptimizer):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
186
187

    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
188
                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
189
190
                 bf16, grad_scaler,
                 models):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
191

Lawrence McAfee's avatar
Lawrence McAfee committed
192
        super().__init__(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
193
            optimizer, clip_grad, log_num_zeros_in_grad,
194
            params_have_main_grad, use_contiguous_buffers_in_local_ddp)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
195

196
197
198
        # >>>
        self.models = models
        # <<<
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
199
        self.bf16 = bf16
mohammad's avatar
mohammad committed
200
        self.grad_scaler = grad_scaler
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
201
202
203
        # 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
204
205
206

        # Tensor used to determine if a nan/if has happend.
        # Any non-zero value indicates inf/nan.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
207
208
209
210
        # 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
211
212

        # Dummy tensor needed for apex multi-apply tensor.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
213
214
215
216
217
218
219
220
221
222
        # 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
223

Lawrence McAfee's avatar
Lawrence McAfee committed
224
225
226
227
228
229
230

    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
231
232
233
234
    def reload_model_params(self):
        self._copy_model_params_to_main_params()


Lawrence McAfee's avatar
Lawrence McAfee committed
235
236
237
238
239
240
241
242
243
244
245
246
247
248
    def _unscale_main_grads_and_check_for_nan(self):

        # 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.
249
250
251
252
253
        # >>>
        # torch.distributed.all_reduce(self.found_inf,
        #                              op=torch.distributed.ReduceOp.MAX,
        #                              group=mpu.get_model_parallel_group())
        # +++
Lawrence McAfee's avatar
Lawrence McAfee committed
254
        torch.distributed.all_reduce(self.found_inf,
255
256
                                     op=torch.distributed.ReduceOp.MAX)
        # <<<
Lawrence McAfee's avatar
Lawrence McAfee committed
257
258
259
260
261
262

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

        return found_inf_flag

Lawrence McAfee's avatar
Lawrence McAfee committed
263
264
    # >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
    @classmethod
265
    def debug_base(cls, ITERATION, key, value):
Lawrence McAfee's avatar
Lawrence McAfee committed
266
267
268
269
270
271
272
        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:
273
                print("            + br/%s; [r%d, i%d]; %s, %.12e" % ("fix " if args.use_distributed_optimizer else "main", my_rank, ITERATION, key, value))
Lawrence McAfee's avatar
Lawrence McAfee committed
274
275
276
277
278
279
280
            torch.distributed.barrier()
        torch.distributed.barrier()
        # if my_rank == 0:
        #     raise Exception("debug.")
        # else:
        #     exit(0)
        exit(0)
281
282
    def debug_model(self, ITERATION, key, use_grad):
        use_grad = bool(use_grad)
283
        tensors = [
284
            (p.main_grad.float() if use_grad else p.float())
285
286
287
            for m in self.models for p in m.parameters()
        ]
        count = sum(t.nelement() for t in tensors)
288
        return self.debug_base(
289
290
            ITERATION,
            "model/%s, %s [count %d]" % (
291
                "grad" if use_grad else "param",
292
293
294
                key,
                count,
            ),
295
296
            # sum(torch.sum(torch.abs(t)) for t in tensors).item() / count,
            sum(torch.sum(torch.abs(t)) for t in tensors),
297
        )
298
299
300
301
302
303
304
305
306
    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)
307
        return self.debug_base(
Lawrence McAfee's avatar
Lawrence McAfee committed
308
            ITERATION,
309
310
311
312
313
314
            "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
315
316
        )
    # <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
Lawrence McAfee's avatar
Lawrence McAfee committed
317
318

    @torch.no_grad()
319
    def step(self, ITERATION):
Lawrence McAfee's avatar
Lawrence McAfee committed
320
321
322

        timers = get_timers()

323
324
325
        # >>>
        # self.debug_model_param(ITERATION, "before copy grad.")
        # self.debug_model_grad(ITERATION, "before copy grad.")
326
327
        # self.debug_main_param(ITERATION, "before copy grad.")
        # self.debug_main_grad(ITERATION, "before copy grad.")
328
329
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
330
331
        # Copy gradients from model params to main params.
        timers('optimizer-copy-to-main-grad').start()
332
        self._copy_model_grads_to_main_grads(ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
333
334
        timers('optimizer-copy-to-main-grad').stop()

335
        # >>>
336
337
        # self.debug_model(ITERATION, "after copy grad.", 0)
        # self.debug_main(ITERATION, "after copy grad.", 1)
338
339
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
        # 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()

            # 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:
355
356
357
358
359
                pax(0, {
                    "main params" : self.get_main_params(),
                    "main grads" : self.get_main_grads(),
                    "found_inf_flag" : found_inf_flag,
                })
Lawrence McAfee's avatar
Lawrence McAfee committed
360
361
362
363
364
365
                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
366
            grad_norm = self.clip_grad_norm(self.clip_grad, ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
367
368
369
370
371
372
        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

373
374
375
376
377
378
379
380
381
382
383
384
        # >>>
        # param = self.optimizer.param_groups[0]["params"][0]
        # pax(0, {
        #     "param" : tp(param),
        #     "grad" : tp(param.grad),
        # })
        # <<<

        # >>>
        # self.debug_main(ITERATION, "before step.", 0)
        # <<<

385
386
387
        # Step the optimizer.
        self.optimizer.step()

Lawrence McAfee's avatar
Lawrence McAfee committed
388
        # >>>
389
        # self.debug_main(ITERATION, "after step.", 0)
Lawrence McAfee's avatar
Lawrence McAfee committed
390
391
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
392
393
        # Update params from main params.
        timers('optimizer-copy-main-to-model-params').start()
394
        self._copy_main_params_to_model_params(ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
395
396
        timers('optimizer-copy-main-to-model-params').stop()

397
        # >>>
398
399
        # self.debug_main_param(ITERATION, "after copy param.")
        # self.debug_main_grad(ITERATION, "after copy param.")
400
401
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
402
403
404
405
        # Successful update.
        return True, grad_norm, num_zeros_in_grad


Lawrence McAfee's avatar
Lawrence McAfee committed
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
# class Float16OptimizerWithFloat16Params(MegatronOptimizer):
class Float16OptimizerWithFloat16Params(BaseFloat16Optimizer):
    """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
434
                 bf16, grad_scaler, models):
Lawrence McAfee's avatar
Lawrence McAfee committed
435
436
437
438

        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
439
            bf16, grad_scaler, models)
Lawrence McAfee's avatar
Lawrence McAfee committed
440

mohammad's avatar
mohammad committed
441
        # ======================
442
        # main parameter stuff
mohammad's avatar
mohammad committed
443
444
445
        # ======================

        # Three groups of parameters:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
446
447
        #   float16_groups: original float16 parameters
        #   fp32_from_float16_groups: fp32 copy of float16 parameters
mohammad's avatar
mohammad committed
448
        #   fp32_from_fp32_groups: original fp32 parameters
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
449
450
        self.float16_groups = []
        self.fp32_from_float16_groups = []
mohammad's avatar
mohammad committed
451
452
453
454
        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
455
            float16_params_this_group = []
mohammad's avatar
mohammad committed
456
            fp32_params_this_group = []
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
457
            fp32_from_float16_params_this_group = []
mohammad's avatar
mohammad committed
458
459
460
461
            # 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
462
463
464
465
                    # float16 params:
                    if param.type() in ['torch.cuda.HalfTensor',
                                        'torch.cuda.BFloat16Tensor']:
                        float16_params_this_group.append(param)
mohammad's avatar
mohammad committed
466
                        # Create a copy
467
                        main_param = param.detach().clone().float()
mohammad's avatar
mohammad committed
468
                        # Copy tensor model parallel attributes.
469
                        mpu.copy_tensor_model_parallel_attributes(main_param,
mohammad's avatar
mohammad committed
470
                                                                  param)
471
                        if hasattr(param, 'shared'):
472
                            main_param.shared = param.shared
mohammad's avatar
mohammad committed
473
                        # Replace the optimizer params with the new fp32 copy.
474
                        param_group['params'][i] = main_param
475

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
476
                        fp32_from_float16_params_this_group.append(main_param)
477
                        # Reset existing state dict key to the new main param.
mohammad's avatar
mohammad committed
478
                        if param in self.optimizer.state:
479
480
481
                            # >>>
                            raise Exception("hi.")
                            # <<<
482
                            self.optimizer.state[main_param] \
mohammad's avatar
mohammad committed
483
484
485
486
                                = self.optimizer.state.pop(param)

                    # fp32 params.
                    elif param.type() == 'torch.cuda.FloatTensor':
Lawrence McAfee's avatar
Lawrence McAfee committed
487
488
489
                        # >>>
                        pax(0, {"param": param})
                        # <<<
mohammad's avatar
mohammad committed
490
491
492
493
                        fp32_params_this_group.append(param)
                        param_group['params'][i] = param

                    else:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
494
495
496
497
498
499
500
501
502
                        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
503
504
505
506
507
508
            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())

Lawrence McAfee's avatar
Lawrence McAfee committed
509
510
511
512
513
514
515
516
517
518
        # >>>
        # from megatron.mpu.layers import param_is_not_tensor_parallel_duplicate
        # params = self.get_parameters()
        # pax(0, {
        #     # "params / 0" : params[0],
        #     "params" : [ (p.tensor_model_parallel, tp(p)) for p in params ],
        #     "grads" : [ (param_is_not_tensor_parallel_duplicate(p.grad), tp(p.grad)) for p in params ],
        # })
        # <<<

mohammad's avatar
mohammad committed
519
520
521

    def zero_grad(self, set_to_none=True):
        """We only need to zero the model related parameters, i.e.,
522
523
524
525
        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
526
        for group in self.float16_groups:
mohammad's avatar
mohammad committed
527
            _zero_grad_group_helper(group, set_to_none)
528
529
        for group in self.fp32_from_float16_groups:
            _zero_grad_group_helper(group, set_to_none)
mohammad's avatar
mohammad committed
530
531
532
533
        for group in self.fp32_from_fp32_groups:
            _zero_grad_group_helper(group, set_to_none)


534
    # >>>
535
    def reduce_grads(self, model):
536
537
538
539
540
541
542
543
544
545
546
547
548

        # >>>
        from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP

        from megatron import get_args
        from megatron import get_timers
        from megatron.model import DistributedDataParallel as LocalDDP
        from megatron.model import Float16Module
        from megatron.utils import unwrap_model

        args = get_args()
        timers = get_timers()
        # <<<
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563

        # All-reduce if needed.
        if args.DDP_impl == 'local':
            timers('backward-params-all-reduce').start()
            for model_module in model:
                model_module.allreduce_gradients()
            timers('backward-params-all-reduce').stop()

        # 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).
        timers('backward-embedding-all-reduce').start()
        if mpu.is_rank_in_embedding_group(ignore_virtual=True) and \
                mpu.get_pipeline_model_parallel_world_size() > 1:
564
            # >>>
565
            # raise Exception("[main] ready for weight sync?")
566
            # <<<
567
568
569
570
571
572
573
574
575
576
577
            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]
            else:  # We do not support the interleaved schedule for T5 yet.
                unwrapped_model = model[0]
            unwrapped_model = unwrap_model(
                unwrapped_model, (torchDDP, LocalDDP, Float16Module))

            if unwrapped_model.share_word_embeddings:
                word_embeddings_weight = unwrapped_model.word_embeddings_weight()
578
579
580
581
582
                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())
583
584
585
586
587
588
589

        # 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
        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:
590
591
592
            # >>>
            raise Exception("[main] ready for t5 sync?")
            # <<<
593
594
595
596
597
            unwrapped_model = model[0]
            unwrapped_model = unwrap_model(
                unwrapped_model, (torchDDP, LocalDDP, Float16Module))
            assert args.DDP_impl == 'local', \
                'T5 model is only supported with local DDP mode'
598
599
            grad = unwrapped_model.language_model.embedding.position_embeddings.weight.main_grad
            torch.distributed.all_reduce(grad, group=mpu.get_position_embedding_group())
600
601
        timers('backward-embedding-all-reduce').stop()

602
    def gather_params(self, ITERATION):
Lawrence McAfee's avatar
Lawrence McAfee committed
603
        pass
Lawrence McAfee's avatar
Lawrence McAfee committed
604

605
    def _copy_model_grads_to_main_grads(self, ITERATION):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
606
607
608
        # 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):
609
            for model_param, main_param in zip(model_group, main_group):
610
                if self.params_have_main_grad and hasattr(model_param, 'main_grad'):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
611
612
613
614
                    main_param.grad = model_param.main_grad.float()
                else:
                    if model_param.grad is not None:
                        main_param.grad = model_param.grad.float()
615
616
617
618
619

                # 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
620
                if self.params_have_main_grad and \
621
                   not self.use_contiguous_buffers_in_local_ddp:
622
623
                    model_param.main_grad = None

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
624
625
626
627
628
        # 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
629

630
631
632
                    # 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.)
633
                    if not self.use_contiguous_buffers_in_local_ddp:
634
                        model_param.main_grad = None
mohammad's avatar
mohammad committed
635

636
637
638
639
640
641
642
643
644
645
        # >>>
        # if ITERATION == DEBUG_ITERATION:
        #     pax(0, {
        #         "** branch **" : "** main. **",
        #         "ITERATION" : ITERATION,
        #         "model grads" :
        #         [ p.main_grad for m in self.models for p in m.parameters() ],
        #     })
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
646
647
    def _collect_main_grad_data_for_unscaling(self):

648
        main_grads = []
Lawrence McAfee's avatar
Lawrence McAfee committed
649
650

        # fp32 params from float16 ones.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
651
        for main_group in self.fp32_from_float16_groups:
652
653
654
            for main_param in main_group:
                if main_param.grad is not None:
                    main_grads.append(main_param.grad.data)
Lawrence McAfee's avatar
Lawrence McAfee committed
655

mohammad's avatar
mohammad committed
656
        # Append fp32 parameters.
657
658
659
660
        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)
Lawrence McAfee's avatar
Lawrence McAfee committed
661
662
        
        return main_grads
mohammad's avatar
mohammad committed
663
664


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
665
    def _get_model_and_main_params_data_float16(self):
mohammad's avatar
mohammad committed
666
        model_data = []
667
        main_data = []
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
668
669
        for model_group, main_group in zip(self.float16_groups,
                                           self.fp32_from_float16_groups):
670
            for model_param, main_param in zip(model_group, main_group):
mohammad's avatar
mohammad committed
671
                model_data.append(model_param.data)
672
673
                main_data.append(main_param.data)
        return model_data, main_data
674
675


676
    def _copy_main_params_to_model_params(self, ITERATION):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
677
678
        # Only needed for the float16 params.
        model_data, main_data = self._get_model_and_main_params_data_float16()
679
680
        _multi_tensor_copy_this_to_that(this=main_data, that=model_data,
                                        overflow_buf=self._dummy_overflow_buf)
681
        # >>>
Lawrence McAfee's avatar
Lawrence McAfee committed
682
683
684
685
686
687
        # if ITERATION == DEBUG_ITERATION:
        #     pax(0, {
        #         "** branch **" : "** main. **",
        #         "ITERATION" : ITERATION,
        #         "model params" : [p for m in self.models for p in m.parameters()],
        #     })
688
        # <<<
689
690
691


    def _copy_model_params_to_main_params(self):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
692
693
        # Only needed for the float16 params.
        model_data, main_data = self._get_model_and_main_params_data_float16()
694
695
        _multi_tensor_copy_this_to_that(this=model_data, that=main_data,
                                        overflow_buf=self._dummy_overflow_buf)
696
697


mohammad's avatar
mohammad committed
698
699
700
    def state_dict(self):
        state_dict = {}
        state_dict['optimizer'] = self.optimizer.state_dict()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
701
702
703
        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
704
705
706
707
        return state_dict


    def load_state_dict(self, state_dict):
mohammad's avatar
mohammad committed
708
709
710
711
712
713
714
715
716
717
718
719
720
        # 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
721
722
723
724
725
726
            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
727

728
        # Copy data for the main params.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
729
730
731
        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
732
        for current_group, saved_group in zip(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
733
734
                self.fp32_from_float16_groups,
                state_dict[fp32_from_float16_params_key]):
mohammad's avatar
mohammad committed
735
736
737
738
            for current_param, saved_param in zip(current_group, saved_group):
                current_param.data.copy_(saved_param.data)


mohammad's avatar
mohammad committed
739
740
class FP32Optimizer(MegatronOptimizer):

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
741
742
    def __init__(self, optimizer, clip_grad,
                 log_num_zeros_in_grad,
743
                 params_have_main_grad,
744
                 use_contiguous_buffers_in_local_ddp):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
745
746
747

        super(FP32Optimizer, self).__init__(
            optimizer, clip_grad, log_num_zeros_in_grad,
748
            params_have_main_grad, use_contiguous_buffers_in_local_ddp)
mohammad's avatar
mohammad committed
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766

        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()
    def step(self):
        """Clip gradients (if needed) and step the base optimizer.
mohammad's avatar
mohammad committed
767
        Always return successful since there is no overflow."""
mohammad's avatar
mohammad committed
768

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
769
770
771
772
773
774
        # 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

775
776
777
                    # 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.)
778
                    if not self.use_contiguous_buffers_in_local_ddp:
779
780
                        param.main_grad = None

mohammad's avatar
mohammad committed
781
        # Clip gradients.
782
        grad_norm = None
mohammad's avatar
mohammad committed
783
        if self.clip_grad > 0.0:
784
            grad_norm = self.clip_grad_norm(self.clip_grad)
mohammad's avatar
mohammad committed
785

Rewon Child's avatar
Rewon Child committed
786
        # count the zeros in the grads
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
787
788
        num_zeros_in_grad = self.count_zeros() if \
                            self.log_num_zeros_in_grad else None
Rewon Child's avatar
Rewon Child committed
789

mohammad's avatar
mohammad committed
790
791
792
793
        # Update parameters.
        self.optimizer.step()

        # No overflow for FP32 optimizer.
794
        return True, grad_norm, num_zeros_in_grad
mohammad's avatar
mohammad committed
795
796


797
798
799
800
    def reload_model_params(self):
        pass


mohammad's avatar
mohammad committed
801
802
803
804
805
806
    def state_dict(self):
        return self.optimizer.state_dict()


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