"vscode:/vscode.git/clone" did not exist on "8116804e4f6e1bf8b205683b813a71ecbc30170b"
optimizer.py 61.4 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
    def clip_grad_norm(self, clip_grad, ITERATION):
Lawrence McAfee's avatar
Lawrence McAfee committed
102
        # >>>
Lawrence McAfee's avatar
Lawrence McAfee committed
103
        return
Lawrence McAfee's avatar
Lawrence McAfee committed
104
        # <<<
Lawrence McAfee's avatar
Lawrence McAfee committed
105
106
        params = self.get_parameters()
        return clip_grad_norm_fp32(params, clip_grad, ITERATION = ITERATION)
107

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
108

Rewon Child's avatar
Rewon Child committed
109
110
111
112
    def count_zeros(self):
        params = self.get_parameters()
        return count_zeros_fp32(params)

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
113

mohammad's avatar
mohammad committed
114
115
116
117
    @abstractmethod
    def zero_grad(self, set_to_none=True):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
118

mohammad's avatar
mohammad committed
119
120
    @abstractmethod
    def get_loss_scale(self):
121
        """The output should be a cuda tensor of size 1."""
mohammad's avatar
mohammad committed
122
123
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
124

mohammad's avatar
mohammad committed
125
126
127
128
    def scale_loss(self, loss):
        """Simple scaling."""
        return self.get_loss_scale() * loss

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
129

Lawrence McAfee's avatar
Lawrence McAfee committed
130
    @abstractmethod
131
    def reduce_grads(self):
Lawrence McAfee's avatar
Lawrence McAfee committed
132
133
134
        pass


mohammad's avatar
mohammad committed
135
136
137
138
    @abstractmethod
    def step(self):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
139

Lawrence McAfee's avatar
Lawrence McAfee committed
140
141
142
143
144
    @abstractmethod
    def gather_params(self):
        pass


145
146
    @abstractmethod
    def reload_model_params(self):
147
148
149
150
151
        """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."""
152
153
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
154

mohammad's avatar
mohammad committed
155
156
157
158
    @abstractmethod
    def state_dict(self):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
159

mohammad's avatar
mohammad committed
160
161
162
163
    @abstractmethod
    def load_state_dict(self, state_dict):
        pass

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
164

mohammad's avatar
mohammad committed
165
166
167
168
169
170
171
172
173
174
    # 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
175

mohammad's avatar
mohammad committed
176
177
178
179
180
181
182
183
184
185
186
187
    # 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
188
class BaseFloat16Optimizer(MegatronOptimizer):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
189
190

    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
191
                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
192
193
                 bf16, grad_scaler,
                 models):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
194

Lawrence McAfee's avatar
Lawrence McAfee committed
195
        super().__init__(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
196
            optimizer, clip_grad, log_num_zeros_in_grad,
197
            params_have_main_grad, use_contiguous_buffers_in_local_ddp)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
198

199
200
201
        # >>>
        self.models = models
        # <<<
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
202
        self.bf16 = bf16
mohammad's avatar
mohammad committed
203
        self.grad_scaler = grad_scaler
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
204
205
206
        # 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
207
208
209

        # Tensor used to determine if a nan/if has happend.
        # Any non-zero value indicates inf/nan.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
210
211
212
213
        # 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
214
215

        # Dummy tensor needed for apex multi-apply tensor.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
216
217
218
219
220
221
222
223
224
225
        # 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
226

Lawrence McAfee's avatar
Lawrence McAfee committed
227
228
229
230
231
232
233

    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
234
235
236
237
    def reload_model_params(self):
        self._copy_model_params_to_main_params()


Lawrence McAfee's avatar
Lawrence McAfee committed
238
239
240
241
242
243
244
245
246
247
248
249
250
251
    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.
252
253
254
255
256
        # >>>
        # 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
257
        torch.distributed.all_reduce(self.found_inf,
258
259
                                     op=torch.distributed.ReduceOp.MAX)
        # <<<
Lawrence McAfee's avatar
Lawrence McAfee committed
260
261
262
263
264
265

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

        return found_inf_flag

Lawrence McAfee's avatar
Lawrence McAfee committed
266
267
268
269
270
271
272
273
274
275
    # >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
    @classmethod
    def debug_general(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:
276
                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
277
278
279
280
281
282
283
284
            torch.distributed.barrier()
        torch.distributed.barrier()
        # if my_rank == 0:
        #     raise Exception("debug.")
        # else:
        #     exit(0)
        exit(0)

285
286
287
    # def _debug_model(self, ITERATION, key, use_param):
    def debug_model(self, ITERATION, key, use_grad):
        use_grad = bool(use_grad)
288
        tensors = [
289
            (p.main_grad.float() if use_grad else p.float())
290
291
292
293
294
295
296
297
298
299
300
            for m in self.models for p in m.parameters()
        ]
        # pax(0, {
        #     "params" : params,
        #     "params / abs" : [ torch.abs(p) for p in params ],
        #     "params / abs / sum" : [ torch.sum(torch.abs(p)) for p in params ],
        # })
        count = sum(t.nelement() for t in tensors)
        return self.debug_general(
            ITERATION,
            "model/%s, %s [count %d]" % (
301
                "grad" if use_grad else "param",
302
303
304
                key,
                count,
            ),
305
306
            # sum(torch.sum(torch.abs(t)) for t in tensors).item() / count,
            sum(torch.sum(torch.abs(t)) for t in tensors),
307
        )
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
    # def debug_model_param(self, ITERATION, key):
    #     return self._debug_model(ITERATION, key, True)
    # def debug_model_grad(self, ITERATION, key):
    #     return self._debug_model(ITERATION, key, False)

    # def _debug_main(self, ITERATION, key0, key1, f, ff):
    #     count = sum(
    #         p.nelement()
    #         for g in self.optimizer.param_groups
    #         for p in g["params"]
    #     )
    #     return self.debug_general(
    #         ITERATION,
    #         "main/%s, %s [count %d]" % (key1, key0, count),
    #         sum(ff(f(p))
    #             for g in self.optimizer.param_groups
    #             for p in g["params"]).item() / count,
    #     )
    # def debug_main_param(self, ITERATION, key):
Lawrence McAfee's avatar
Lawrence McAfee committed
327
328
329
    #     return self._debug_main(
    #         ITERATION,
    #         key,
330
331
332
333
    #         "param", # sum",
    #         # lambda p : p,
    #         lambda p : torch.abs(p),
    #         torch.sum,
Lawrence McAfee's avatar
Lawrence McAfee committed
334
    #     )
335
    # def debug_main_grad(self, ITERATION, key):
Lawrence McAfee's avatar
Lawrence McAfee committed
336
337
338
    #     return self._debug_main(
    #         ITERATION,
    #         key,
339
340
341
342
    #         "grad", # sum",
    #         # lambda p : p.grad,
    #         lambda p : torch.abs(p.grad),
    #         torch.sum,
Lawrence McAfee's avatar
Lawrence McAfee committed
343
    #     )
344
345
346
347
348
349
350
351
352
353
354
    # def _debug_main(self, ITERATION, key, use_param):
    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_general(
Lawrence McAfee's avatar
Lawrence McAfee committed
355
            ITERATION,
356
357
358
359
360
361
            "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
362
        )
363
364
365
366
    # def debug_main_param(self, ITERATION, key):
    #     return self._debug_main(ITERATION, key, True)
    # def debug_main_grad(self, ITERATION, key):
    #     return self._debug_main(ITERATION, key, False)
Lawrence McAfee's avatar
Lawrence McAfee committed
367
    # <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
Lawrence McAfee's avatar
Lawrence McAfee committed
368
369

    @torch.no_grad()
370
    def step(self, ITERATION):
Lawrence McAfee's avatar
Lawrence McAfee committed
371
372
373

        timers = get_timers()

374
375
376
        # >>>
        # self.debug_model_param(ITERATION, "before copy grad.")
        # self.debug_model_grad(ITERATION, "before copy grad.")
377
378
        # self.debug_main_param(ITERATION, "before copy grad.")
        # self.debug_main_grad(ITERATION, "before copy grad.")
379
380
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
381
382
        # Copy gradients from model params to main params.
        timers('optimizer-copy-to-main-grad').start()
383
        self._copy_model_grads_to_main_grads(ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
384
385
        timers('optimizer-copy-to-main-grad').stop()

386
        # >>>
387
388
        # self.debug_model(ITERATION, "after copy grad.", 0)
        # self.debug_main(ITERATION, "after copy grad.", 1)
389
390
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
        # 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:
406
407
408
409
410
                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
411
412
413
414
415
416
                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
417
            grad_norm = self.clip_grad_norm(self.clip_grad, ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
418
419
420
421
422
423
        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

424
425
426
427
428
429
430
431
432
433
434
435
        # >>>
        # param = self.optimizer.param_groups[0]["params"][0]
        # pax(0, {
        #     "param" : tp(param),
        #     "grad" : tp(param.grad),
        # })
        # <<<

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

436
437
438
        # Step the optimizer.
        self.optimizer.step()

Lawrence McAfee's avatar
Lawrence McAfee committed
439
        # >>>
440
        # self.debug_main(ITERATION, "after step.", 0)
Lawrence McAfee's avatar
Lawrence McAfee committed
441
442
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
443
444
        # Update params from main params.
        timers('optimizer-copy-main-to-model-params').start()
445
        self._copy_main_params_to_model_params(ITERATION)
Lawrence McAfee's avatar
Lawrence McAfee committed
446
447
        timers('optimizer-copy-main-to-model-params').stop()

448
        # >>>
449
450
        # self.debug_main_param(ITERATION, "after copy param.")
        # self.debug_main_grad(ITERATION, "after copy param.")
451
452
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
453
454
455
456
        # Successful update.
        return True, grad_norm, num_zeros_in_grad


Lawrence McAfee's avatar
Lawrence McAfee committed
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
# 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
485
                 bf16, grad_scaler, models):
Lawrence McAfee's avatar
Lawrence McAfee committed
486
487
488
489

        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
490
            bf16, grad_scaler, models)
Lawrence McAfee's avatar
Lawrence McAfee committed
491

mohammad's avatar
mohammad committed
492
        # ======================
493
        # main parameter stuff
mohammad's avatar
mohammad committed
494
495
496
        # ======================

        # Three groups of parameters:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
497
498
        #   float16_groups: original float16 parameters
        #   fp32_from_float16_groups: fp32 copy of float16 parameters
mohammad's avatar
mohammad committed
499
        #   fp32_from_fp32_groups: original fp32 parameters
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
500
501
        self.float16_groups = []
        self.fp32_from_float16_groups = []
mohammad's avatar
mohammad committed
502
503
504
505
        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
506
            float16_params_this_group = []
mohammad's avatar
mohammad committed
507
            fp32_params_this_group = []
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
508
            fp32_from_float16_params_this_group = []
mohammad's avatar
mohammad committed
509
510
511
512
            # 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
513
514
515
516
                    # float16 params:
                    if param.type() in ['torch.cuda.HalfTensor',
                                        'torch.cuda.BFloat16Tensor']:
                        float16_params_this_group.append(param)
mohammad's avatar
mohammad committed
517
                        # Create a copy
518
                        main_param = param.detach().clone().float()
mohammad's avatar
mohammad committed
519
                        # Copy tensor model parallel attributes.
520
                        mpu.copy_tensor_model_parallel_attributes(main_param,
mohammad's avatar
mohammad committed
521
                                                                  param)
522
                        if hasattr(param, 'shared'):
523
                            main_param.shared = param.shared
mohammad's avatar
mohammad committed
524
                        # Replace the optimizer params with the new fp32 copy.
525
                        param_group['params'][i] = main_param
526

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
527
                        fp32_from_float16_params_this_group.append(main_param)
528
                        # Reset existing state dict key to the new main param.
mohammad's avatar
mohammad committed
529
                        if param in self.optimizer.state:
530
531
532
                            # >>>
                            raise Exception("hi.")
                            # <<<
533
                            self.optimizer.state[main_param] \
mohammad's avatar
mohammad committed
534
535
536
537
                                = self.optimizer.state.pop(param)

                    # fp32 params.
                    elif param.type() == 'torch.cuda.FloatTensor':
Lawrence McAfee's avatar
Lawrence McAfee committed
538
539
540
                        # >>>
                        pax(0, {"param": param})
                        # <<<
mohammad's avatar
mohammad committed
541
542
543
544
                        fp32_params_this_group.append(param)
                        param_group['params'][i] = param

                    else:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
545
546
547
548
549
550
551
552
553
                        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
554
555
556
557
558
559
            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
560
561
562
563
564
565
566
567
568
569
        # >>>
        # 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
570
571
572

    def zero_grad(self, set_to_none=True):
        """We only need to zero the model related parameters, i.e.,
573
574
575
576
        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
577
        for group in self.float16_groups:
mohammad's avatar
mohammad committed
578
            _zero_grad_group_helper(group, set_to_none)
579
580
        for group in self.fp32_from_float16_groups:
            _zero_grad_group_helper(group, set_to_none)
mohammad's avatar
mohammad committed
581
582
583
584
        for group in self.fp32_from_fp32_groups:
            _zero_grad_group_helper(group, set_to_none)


585
    # >>>
586
    def reduce_grads(self, model):
587
588
589
590
591
592
593
594
595
596
597
598
599

        # >>>
        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()
        # <<<
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614

        # 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:
615
            # >>>
616
            # raise Exception("[main] ready for weight sync?")
617
            # <<<
618
619
620
621
622
623
624
625
626
627
628
            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()
629
630
631
632
633
                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())
634
635
636
637
638
639
640

        # 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:
641
642
643
            # >>>
            raise Exception("[main] ready for t5 sync?")
            # <<<
644
645
646
647
648
            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'
649
650
            grad = unwrapped_model.language_model.embedding.position_embeddings.weight.main_grad
            torch.distributed.all_reduce(grad, group=mpu.get_position_embedding_group())
651
652
        timers('backward-embedding-all-reduce').stop()

653
    def gather_params(self, ITERATION):
Lawrence McAfee's avatar
Lawrence McAfee committed
654
        pass
Lawrence McAfee's avatar
Lawrence McAfee committed
655

656
    def _copy_model_grads_to_main_grads(self, ITERATION):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
657
658
659
        # 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):
660
            for model_param, main_param in zip(model_group, main_group):
661
                if self.params_have_main_grad and hasattr(model_param, 'main_grad'):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
662
663
664
665
                    main_param.grad = model_param.main_grad.float()
                else:
                    if model_param.grad is not None:
                        main_param.grad = model_param.grad.float()
666
667
668
669
670

                # 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
671
                if self.params_have_main_grad and \
672
                   not self.use_contiguous_buffers_in_local_ddp:
673
674
                    model_param.main_grad = None

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
675
676
677
678
679
        # 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
680

681
682
683
                    # 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.)
684
                    if not self.use_contiguous_buffers_in_local_ddp:
685
                        model_param.main_grad = None
mohammad's avatar
mohammad committed
686

687
688
689
690
691
692
693
694
695
696
        # >>>
        # 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
697
698
    def _collect_main_grad_data_for_unscaling(self):

699
        main_grads = []
Lawrence McAfee's avatar
Lawrence McAfee committed
700
701

        # fp32 params from float16 ones.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
702
        for main_group in self.fp32_from_float16_groups:
703
704
705
            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
706

mohammad's avatar
mohammad committed
707
        # Append fp32 parameters.
708
709
710
711
        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
712
713
        
        return main_grads
mohammad's avatar
mohammad committed
714
715


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
716
    def _get_model_and_main_params_data_float16(self):
mohammad's avatar
mohammad committed
717
        model_data = []
718
        main_data = []
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
719
720
        for model_group, main_group in zip(self.float16_groups,
                                           self.fp32_from_float16_groups):
721
            for model_param, main_param in zip(model_group, main_group):
mohammad's avatar
mohammad committed
722
                model_data.append(model_param.data)
723
724
                main_data.append(main_param.data)
        return model_data, main_data
725
726


727
    def _copy_main_params_to_model_params(self, ITERATION):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
728
729
        # Only needed for the float16 params.
        model_data, main_data = self._get_model_and_main_params_data_float16()
730
731
        _multi_tensor_copy_this_to_that(this=main_data, that=model_data,
                                        overflow_buf=self._dummy_overflow_buf)
732
        # >>>
Lawrence McAfee's avatar
Lawrence McAfee committed
733
734
735
736
737
738
        # if ITERATION == DEBUG_ITERATION:
        #     pax(0, {
        #         "** branch **" : "** main. **",
        #         "ITERATION" : ITERATION,
        #         "model params" : [p for m in self.models for p in m.parameters()],
        #     })
739
        # <<<
740
741
742


    def _copy_model_params_to_main_params(self):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
743
744
        # Only needed for the float16 params.
        model_data, main_data = self._get_model_and_main_params_data_float16()
745
746
        _multi_tensor_copy_this_to_that(this=model_data, that=main_data,
                                        overflow_buf=self._dummy_overflow_buf)
747
748


mohammad's avatar
mohammad committed
749
750
751
    def state_dict(self):
        state_dict = {}
        state_dict['optimizer'] = self.optimizer.state_dict()
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
752
753
754
        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
755
756
757
758
        return state_dict


    def load_state_dict(self, state_dict):
mohammad's avatar
mohammad committed
759
760
761
762
763
764
765
766
767
768
769
770
771
        # 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
772
773
774
775
776
777
            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
778

779
        # Copy data for the main params.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
780
781
782
        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
783
        for current_group, saved_group in zip(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
784
785
                self.fp32_from_float16_groups,
                state_dict[fp32_from_float16_params_key]):
mohammad's avatar
mohammad committed
786
787
788
789
            for current_param, saved_param in zip(current_group, saved_group):
                current_param.data.copy_(saved_param.data)


790
# >>>
791
import math
792

793
from megatron import get_args
794

Lawrence McAfee's avatar
Lawrence McAfee committed
795
796
797
798
799
800
801
802
803
804
805
# class ShardIndex:
class Shard:
    def __init__(self, start, end):
        self.start = start
        self.end = end
        self.size = end - start
    def normalize(self, start = 0):
        return Shard(start, start + self.size)
    def __str__(self):
        return "%d,%d [%d]" % (self.start, self.end, self.size)

806
# class Float16DistributedOptimizer(Float16OptimizerWithFloat16Params):
Lawrence McAfee's avatar
Lawrence McAfee committed
807
808
# class Float16DistributedOptimizer(MegatronOptimizer):
class Float16DistributedOptimizer(BaseFloat16Optimizer):
809

810
    @classmethod
Lawrence McAfee's avatar
Lawrence McAfee committed
811
    def get_model_gbuf_param_shard_map(cls, model, dtype, gbuf_world_shard):
812

Lawrence McAfee's avatar
Lawrence McAfee committed
813
814
        # Param shard map.
        param_world_index_map = model._grad_buffer_param_index_map[dtype]
815
        param_shard_map = {}
Lawrence McAfee's avatar
Lawrence McAfee committed
816
        for param, param_world_indexes in param_world_index_map.items():
817

Lawrence McAfee's avatar
Lawrence McAfee committed
818
819
820
            # Shard range.
            param_world_start, param_world_end = param_world_indexes
            param_local_start = max(
821
                0,
Lawrence McAfee's avatar
Lawrence McAfee committed
822
823
824
825
826
827
828
829
                param_world_start - gbuf_world_shard.start)
            param_local_end = min(
                gbuf_world_shard.size,
                param_world_end - gbuf_world_shard.start)

            # Add shard, if within range.
            if param_local_end > param_local_start:
                param_local_shard = Shard(param_local_start, param_local_end)
Lawrence McAfee's avatar
Lawrence McAfee committed
830
831
832
                # param_world_shard = param_local_shard.normalize(param_world_start)
                param_world_shard = param_local_shard.normalize(
                    param_local_start + gbuf_world_shard.start)
833
834
                sub_param_start = max(0, gbuf_world_shard.start-param_world_start)
                sub_param_shard = param_local_shard.normalize(sub_param_start)
Lawrence McAfee's avatar
Lawrence McAfee committed
835
                param_shard_map[param] = {
836
837
838
                    "gbuf_world" : param_world_shard,
                    "gbuf_local" : param_local_shard,
                    "param" : sub_param_shard,
839
840
                }

Lawrence McAfee's avatar
Lawrence McAfee committed
841
        # pax(0, {"param_shard_map": [ str((str(p.shape), s)) for p,s in param_shard_map.items() ]})
842
843
844
845

        return param_shard_map

    @classmethod
Lawrence McAfee's avatar
Lawrence McAfee committed
846
    def get_model_gbuf_shard(cls, model, dtype):
847

Lawrence McAfee's avatar
Lawrence McAfee committed
848
849
        data_parallel_rank = mpu.get_data_parallel_rank()
        data_parallel_world_size = mpu.get_data_parallel_world_size()
850
851

        # Grad buffer shard.
Lawrence McAfee's avatar
Lawrence McAfee committed
852
853
854
855
        grad_buffer = model._grad_buffers[dtype]
        gbuf_size = grad_buffer.numel
        max_gbuf_shard_size = int(math.ceil(gbuf_size / data_parallel_world_size))

856
857
858
859
860
861
        gbuf_world_all_shards = []
        for r in range(data_parallel_world_size):
            gbuf_world_start = r * max_gbuf_shard_size
            gbuf_world_end = min(gbuf_size, gbuf_world_start+max_gbuf_shard_size)
            gbuf_world_shard = Shard(gbuf_world_start, gbuf_world_end)
            gbuf_world_all_shards.append(gbuf_world_shard)
862
863
864
865
866
867
868
869
870
871
            # >>>
            # if max_gbuf_shard_size != gbuf_world_shard.size:
            #     raise Exception("%d: smaller, rank %d. [ %d -> %d vs. %d]" % (
            #         data_parallel_rank,
            #         r,
            #         gbuf_size,
            #         max_gbuf_shard_size,
            #         gbuf_world_shard.size,
            #     ))
            # <<<
872
        gbuf_world_shard = gbuf_world_all_shards[data_parallel_rank]
Lawrence McAfee's avatar
Lawrence McAfee committed
873
874
875
876
877
878
879
880
881
882
883
        gbuf_local_shard = gbuf_world_shard.normalize()

        # Param shards.
        param_shard_map = cls.get_model_gbuf_param_shard_map(model,
                                                             dtype,
                                                             gbuf_world_shard)

        # Altogether.
        data = {
            "local" : gbuf_local_shard,
            "world" : gbuf_world_shard,
884
            "world_all" : gbuf_world_all_shards,
Lawrence McAfee's avatar
Lawrence McAfee committed
885
            "param_map" : param_shard_map,
886
            "max_shard_size" : max_gbuf_shard_size,
887
888
        }

889
        # pax(0, {"data": data})
890

Lawrence McAfee's avatar
Lawrence McAfee committed
891
        return data
892
893

    @classmethod
Lawrence McAfee's avatar
Lawrence McAfee committed
894
    def get_model_gbuf_shard_map(cls, model):
895
        return {
Lawrence McAfee's avatar
Lawrence McAfee committed
896
            dtype : cls.get_model_gbuf_shard(model, dtype)
897
898
899
            for dtype in model._grad_buffers
        }

Lawrence McAfee's avatar
Lawrence McAfee committed
900
901
    @classmethod
    def get_param_gbuf_map(cls, model_gbuf_shards):
902

Lawrence McAfee's avatar
Lawrence McAfee committed
903
904
905
906
907
908
909
910
911
912
913
914
915
        param_gbuf_map = {}
        for model_index, model_gbuf_shard_map in enumerate(model_gbuf_shards):
            for dtype, gbuf_shard_map in model_gbuf_shard_map.items():
                for param, param_shard_map in gbuf_shard_map["param_map"].items():
                    # assert param not in param_size_map
                    # param_size_map[param] = param_shard_map["local"].size
                    param_gbuf_map[param] = (model_index, dtype)
                    # pax(0, {
                    #     "dtype" : dtype,
                    #     "gbuf_shard_map" : gbuf_shard_map,
                    #     "param" : tp(param),
                    #     "param_shard_map" : param_shard_map,
                    # })
916

Lawrence McAfee's avatar
Lawrence McAfee committed
917
918
919
920
921
922
        # pax(0, {
        #     "model_gbuf_shards" : model_gbuf_shards,
        #     # "param_size_map" :
        #     # [ (str(p.shape), s) for p, s in param_size_map.items() ],
        #     "param_gbuf_map" : param_gbuf_map,
        # })
923

Lawrence McAfee's avatar
Lawrence McAfee committed
924
        return param_gbuf_map
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945

    @classmethod
    def get_optimizer_group_shards(cls, param_groups, model_gbuf_shards):

        num_groups = len(param_groups)

        # Param group map.
        param_group_map = {}
        for group_index, group in enumerate(param_groups):
            for param in group["params"]:
                assert param.requires_grad
                param_group_map[param] = group_index

        # Optimizer group shards.
        group_shards = [ {"size": 0, "param_map": {}} for _ in param_groups ]
        for model_gbuf_shard_map in model_gbuf_shards:
            for dtype, gbuf_shard_map in model_gbuf_shard_map.items():
                for param in gbuf_shard_map["param_map"]:
                    
                    group_index = param_group_map[param]
                    group_shard = group_shards[group_index]
946
                    param_size = gbuf_shard_map["param_map"][param]["param"].size
947
948
949
950
951

                    param_group_start = group_shard["size"]
                    param_group_end = param_group_start + param_size
                    param_group_shard = Shard(param_group_start, param_group_end)

952
                    # group_shard["max_size"] = gbuf_shard_map["max_shard_size"]
953
954
955
                    group_shard["size"] += param_size
                    group_shard["param_map"][param] = param_group_shard

956
                    # pax(0, {"gbuf_shard_map": gbuf_shard_map})
957
958
959
960
961
962
963
964
965
966
                    # >>>
                    # if torch.distributed.get_rank() == 1:
                    #     print(">>> [%d] ... group %d, size %d, param %s. <<<" % (
                    #         torch.distributed.get_rank(),
                    #         group_index,
                    #         param_size,
                    #         str(tuple(param.shape)),
                    #     ))
                    # <<<

967
968
969
970
971
        # Squeeze zero-size group shards.
        for group_index, group_shard in enumerate(group_shards):
            group_shard["orig_group"] = param_groups[group_index]
        group_shards = [ g for g in group_shards if g["size"] > 0 ]

972
973
        # [ ... x ... ] Synchronize group sizes across ranks.
        
974
        # pax(0, {
975
976
977
978
979
980
        #     "param_group_map": [
        #         (g, str(p.shape))
        #         for p, g in param_group_map.items()
        #     ],
        #     "group_shards" : group_shards,
        # })
981
982
983

        return group_shards

984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
    @classmethod
    def allocate_main_param_shards(cls, opt_group_shards):

        # Allocate main param/grad shard.
        # ** torch.nn.Parameter ??
        # ** MemoryBuffer ??
        allocate_shard = lambda shard_size, dtype : torch.empty(
            (shard_size,),
            dtype = dtype,
            device = torch.cuda.current_device(),
            requires_grad = True)
        
        # main_param_shards = []
        for group_index, group_shard in enumerate(opt_group_shards):

999
1000
1001
1002
            # pax(0, {
            #     "group_shard" : group_shard,
            # })

1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
            group_size = group_shard["size"]
            assert group_size != 0, "temporary check ... remove me."

            # ** todo: for dtype in model_main_dtypes ........ **

            # Allocate shard.
            # if group_size == 0:
            #     main_param = None
            # else:
            main_param = allocate_shard(group_size, torch.float)
            main_param.grad = allocate_shard(group_size, torch.float)
            mpu.set_tensor_model_parallel_attributes(main_param, True, 0, 1)

            # main_param_shards.append(main_param)
            group_shard["orig_group"]["params"] = [ main_param ]

            # # Update optimizer group.
            # self.optimizer.param_groups[group_index]["params"] = [ main_param ]

        # pax(1, {
        #     "opt_group_shards" : opt_group_shards,
        #     "main_param_shards" : main_param_shards,
        # })

        # return main_param_shards

1029
1030
    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
1031
                 bf16, grad_scaler, models):
1032
1033
1034

        super().__init__(
            optimizer, clip_grad, log_num_zeros_in_grad,
Lawrence McAfee's avatar
Lawrence McAfee committed
1035
            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
1036
            bf16, grad_scaler, models)
1037

1038
1039
        # >>>
        args = get_args()
1040
        assert args.use_contiguous_buffers_in_local_ddp # already checked in args
1041
        # <<<
1042

1043
1044
1045
1046
        # Model grad buffer shards.
        self.model_gbuf_shards = []
        for model_index, model in enumerate(self.models):
            self.model_gbuf_shards.append(self.get_model_gbuf_shard_map(model))
Lawrence McAfee's avatar
Lawrence McAfee committed
1047
        self.param_gbuf_map = self.get_param_gbuf_map(self.model_gbuf_shards)
1048

1049
1050
1051
1052
1053
        # Optimizer shards.
        self.opt_group_shards = self.get_optimizer_group_shards(
            self.optimizer.param_groups,
            self.model_gbuf_shards)

1054
1055
        # Allocate main param shards.
        self.allocate_main_param_shards(self.opt_group_shards)
1056

1057
        # >>>
1058
1059
1060
1061
1062
        # pax(0, {
        #     "model_gbuf_shards" : self.model_gbuf_shards,
        #     "opt_group_shards" : self.opt_group_shards,
        #     "main_param_shards" : self.main_param_shards,
        # })
1063
1064
        # <<<

1065
1066
1067
1068
1069
        # Update optimizer groups.
        # - Also, leverage state_dict() and load_state_dict() to
        #   recast preexisting per-param state tensors.
        self.optimizer.param_groups = \
            [ g["orig_group"] for g in self.opt_group_shards ]
Lawrence McAfee's avatar
Lawrence McAfee committed
1070
1071
        self.optimizer.load_state_dict(self.optimizer.state_dict())

1072
1073
1074
1075
1076
1077
        # pax(0, {
        #     # "opt_group_shards" : self.opt_group_shards,
        #     # "param_groups" : self.optimizer.param_groups,
        #     "optimizer" : self.optimizer,
        #     "optimizer / state" : self.optimizer.state,
        # })
1078
        # pax(1, {
1079
1080
1081
1082
1083
        #     "optimizer" : self.optimizer,
        #     **{"optimizer / param_groups / %d" % i : g
        #        for i, g in enumerate(self.optimizer.param_groups)},
        #     "optimizer / state" : self.optimizer.state,
        #     "optimizer / state_dict" : self.optimizer.state_dict(),
1084
1085
1086
1087
        # })

        # Initialize main params.
        self._copy_model_params_to_main_params()
1088

1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
    @staticmethod
    def has_nan_debug(tensors):
        if isinstance(tensors, torch.Tensor):
            tensors = [ tensors ]
        assert isinstance(tensors, list)
        has_nans = [ (not torch.all(torch.isfinite(t)).item()) for t in tensors ]
        has_nan = any(has_nans)
        return has_nan
    def get_local_model_param_views(self):
        '''** FOR DEBUGGING. **'''
        model_param_views = []
        for group_index, opt_group_shard in enumerate(self.opt_group_shards):
            for param, opt_shard in opt_group_shard["param_map"].items():
                model_index, dtype = self.param_gbuf_map[param]
                gbuf_shard_map = \
                    self.model_gbuf_shards[model_index][dtype]["param_map"][param]
                model_param_shard = gbuf_shard_map["param"]
                model_param_views.append(
                    param.view(-1)[model_param_shard.start:model_param_shard.end])
        return model_param_views
    def get_local_model_grad_views(self):
        '''** FOR DEBUGGING. **'''
        model_grad_views = []
        for group_index, opt_group_shard in enumerate(self.opt_group_shards):
            for param, opt_shard in opt_group_shard["param_map"].items():
                model_index, dtype = self.param_gbuf_map[param]
                gbuf = self.models[model_index]._grad_buffers[dtype].data
                gbuf_shard_map = \
                    self.model_gbuf_shards[model_index][dtype]["param_map"][param]
                gbuf_world_shard = gbuf_shard_map["gbuf_world"]
                model_grad_views.append(
                    gbuf[gbuf_world_shard.start:gbuf_world_shard.end])
        return model_grad_views
    def get_world_model_params(self):
        '''** FOR DEBUGGING. **'''
        return [ p for m in self.models for p in m.parameters() ]
1125
1126
1127
    def get_world_model_grads(self):
        '''** FOR DEBUGGING. **'''
        return [ p.main_grad for p in self.get_world_model_params() ]
1128
1129
1130
1131
1132

    def get_main_params(self):
        return [ g["params"][0] for g in self.optimizer.param_groups ]
    def get_main_grads(self):
        return [ p.grad for p in self.get_main_params() ]
1133
    def get_main_param(self, group_index):
1134
1135
        # return self.optimizer.param_groups[group_index]["params"][0]
        return self.get_main_params()[group_index]
1136
1137
1138
    def get_main_grad(self, group_index):
        return self.get_main_param(group_index).grad

1139
1140
1141
1142
1143
1144
    def load_state_dict(self):
        raise Exception("hi.")
    def reload_model_params(self):
        raise Exception("hi.")
    def state_dict(self):
        raise Exception("hi.")
Lawrence McAfee's avatar
Lawrence McAfee committed
1145
1146
1147

    def zero_grad(self, set_to_none=True):

Lawrence McAfee's avatar
Lawrence McAfee committed
1148
        model_params = []
Lawrence McAfee's avatar
Lawrence McAfee committed
1149
1150
        for model in self.models:
            for dtype, param_map in model._grad_buffer_param_index_map.items():
Lawrence McAfee's avatar
Lawrence McAfee committed
1151
1152
1153
1154
                model_params.extend(param_map.keys())
        # main_params = []
        # for main_group in self.optimizer.param_groups:
        #     main_params.extend(main_group["params"])
Lawrence McAfee's avatar
Lawrence McAfee committed
1155

1156
1157
        # ** using contiguous buffer; don't set_to_none **
        _zero_grad_group_helper(model_params, set_to_none = False) # set_to_none)
Lawrence McAfee's avatar
Lawrence McAfee committed
1158
        # _zero_grad_group_helper(params, set_to_none = False)
Lawrence McAfee's avatar
Lawrence McAfee committed
1159

1160
        # pax(0, {"model_params": model_params})
1161

1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
    # def get_model_grad_buffer_dp_views(self):

    #     # >>>
    #     # ** only contiguous grad buffer supported, for now [ TEMPORARY ] **
    #     args = get_args()
    #     assert args.use_contiguous_buffers_in_local_ddp
    #     # <<<

    #     # Grad buffer views.
    #     gbuf_view_items = []
    #     for model_index, model in enumerate(self.models):
    #         for dtype, gbuf_shard in self.model_gbuf_shards[model_index].items():
    #             world_shards = gbuf_shard["world_all"]
    #             gbuf = model._grad_buffers[dtype].data
    #             gbuf_views = [ gbuf[s.start:s.end] for s in world_shards ]
    #             gbuf_view_items.append((model_index, dtype, gbuf_views))

    #             # pax(0, {
    #             #     "world_shards" : world_shards,
    #             #     "gbuf_views" : gbuf_views,
    #             # })

    #     pax(0, {
    #         "gbuf_view_items" : gbuf_view_items,
    #         **{
    #             "views / %d" % i : item[2]
    #             for i, item in enumerate(gbuf_view_items)
    #         },
    #     })

    #     return gbuf_view_items
1193
1194
    def get_model_grad_buffer_dp_views(self):

Lawrence McAfee's avatar
Lawrence McAfee committed
1195
        # >>>
1196
1197
1198
        # ** only contiguous grad buffer supported, for now [ TEMPORARY ] **
        args = get_args()
        assert args.use_contiguous_buffers_in_local_ddp
Lawrence McAfee's avatar
Lawrence McAfee committed
1199
        # <<<
1200

1201
1202
1203
        # data_parallel_rank = mpu.get_data_parallel_rank()
        data_parallel_world_size = mpu.get_data_parallel_world_size()

1204
1205
1206
        # Grad buffer views.
        gbuf_view_items = []
        for model_index, model in enumerate(self.models):
1207
            for dtype, gbuf in model._grad_buffers.items():
1208

1209
1210
1211
                # gbuf_size = gbuf.numel_padded
                assert gbuf.numel_padded % data_parallel_world_size == 0
                shard_size = int(gbuf.numel_padded / data_parallel_world_size)
1212
                # pax(0, {
1213
1214
1215
1216
                #     "numel" : gbuf.numel,
                #     "numel_padded" : gbuf.numel_padded,
                #     "shard_size / f" : gbuf.numel_padded/data_parallel_world_size,
                #     "shard_size / i" : shard_size,
1217
                # })
1218
1219
1220
                gbuf_views = [gbuf.data[(r*shard_size):((r+1)*shard_size)]
                              for r in range(data_parallel_world_size)]
                gbuf_view_items.append((model_index, dtype, gbuf_views))
1221

1222
1223
1224
1225
1226
1227
1228
        # pax(0, {
        #     "gbuf_view_items" : gbuf_view_items,
        #     **{
        #         "views / %d" % i : item[2]
        #         for i, item in enumerate(gbuf_view_items)
        #     },
        # })
1229
1230
1231

        return gbuf_view_items

1232
    def reduce_grads(self, model):
1233

1234
        # >>>
1235
1236
1237
1238
1239
1240
1241
1242
1243
        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()
1244
1245
1246
        timers = get_timers()
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
1247
1248
1249
1250
1251
        # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        # Sync word embedding params.

        # ... todo ...

1252
1253
1254
1255
1256
1257
1258
        # 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:
1259
            # >>>
1260
            # raise Exception("[fix] ready for weight sync?")
1261
            # <<<
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
            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()
                # >>>
                if args.DDP_impl == 'local':
                    grad = word_embeddings_weight.main_grad
                else:
1277
                    raise Exception("only 'main_grad' supported for distrib-opt.")
1278
1279
1280
1281
1282
1283
1284
1285
                    grad = word_embeddings_weight.grad
                torch.distributed.all_reduce(grad, group=mpu.get_embedding_group())
                # +++
                # grad_shard = optimizer.get_grad_shard(word_embeddings)
                # torch.distributed.all_reduce(grad_shard,
                #                              group=mpu.get_embedding_group())
                # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
1286
1287
1288
1289
1290
        # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        # Sync T5 position embedding params.

        # ... todo ...

1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
        # 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:
            # >>>
            raise Exception("[fix] ready for t5 sync?")
            # <<<
            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'
            # >>>
            grad = unwrapped_model.language_model.embedding.position_embeddings.weight.main_grad
            torch.distributed.all_reduce(grad, group=mpu.get_position_embedding_group())
            # +++
            # grad_shard = optimizer.get_grad_shard(
            #     unwrapped_model.language_model.embedding.position_embeddings.weight)
            # torch.distributed.all_reduce(grad_shard,
            #                              group=mpu.get_position_embedding_group())
            # <<<
        timers('backward-embedding-all-reduce').stop()

1316
1317
        # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        # Reduce-scatter.
1318
1319
        # timers('backward-params-reduce-scatter').start()
        timers('backward-params-all-reduce').start()
Lawrence McAfee's avatar
Lawrence McAfee committed
1320
        data_parallel_rank = mpu.get_data_parallel_rank()
1321
        data_parallel_world_size = mpu.get_data_parallel_world_size()
Lawrence McAfee's avatar
Lawrence McAfee committed
1322
        data_parallel_group = mpu.get_data_parallel_group()
1323

1324
        gbuf_view_items = self.get_model_grad_buffer_dp_views()
Lawrence McAfee's avatar
Lawrence McAfee committed
1325

1326
        # pax(0, {"gbuf_views": [g for item in gbuf_view_items for g in item[2]]})
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
        # pax(0, {"gbufs": [
        #     g.data
        #     for m in self.models
        #     for g in m._grad_buffers.values()
        # ]})

        # >>>
        # buffer_.data /= mpu.get_data_parallel_world_size()
        # torch.distributed.all_reduce(
        #     buffer_.data, group=mpu.get_data_parallel_group())
        # <<<
1338

1339
1340
1341
1342
1343
        # >>>
        # self.debug_main_param(0, "before reduce scatter")
        # self.debug_main_grad(0, "before reduce scatter")
        # <<<

1344
        for model_index, dtype, gbuf_views in gbuf_view_items:
1345
1346
            # coalesced /= mpu.get_data_parallel_world_size()
            gbuf = self.models[model_index]._grad_buffers[dtype].data
1347
1348
1349
1350
1351

            # >>>
            # ~~ distributed.py ~~
            # gbuf /= data_parallel_world_size
            # torch.distributed.all_reduce(gbuf, group=data_parallel_group)
1352
1353
1354
            # pax(0, {
            #     "gbuf" : tp(gbuf),
            # })
1355
1356
1357
1358
1359
1360
1361
            # <<<

            # torch.mul(gbuf.data, 1. / data_parallel_world_size, out = gbuf.data)
            # gbuf_views = [ t / data_parallel_world_size for t in gbuf_views ]
            gbuf /= data_parallel_world_size

            # if 1:
1362
1363
            # try:
            # pax(0, {"gbuf_views": gbuf_views})
1364
1365
1366
1367
1368
            torch.distributed.reduce_scatter(
                gbuf_views[data_parallel_rank],
                gbuf_views,
                group = data_parallel_group,
            )
1369
1370
1371
1372
1373
            # except:
            #     pax(0, {
            #         "data_parallel_rank" : data_parallel_rank,
            #         "gbuf_views" : gbuf_views,
            #     })
1374
1375
1376
1377
1378
            # else:
            #     torch.distributed.all_reduce(
            #         gbuf,
            #         group = data_parallel_group,
            #     )
1379
1380
        # timers('backward-params-reduce-scatter').stop()
        timers('backward-params-all-reduce').stop()
1381
            
1382
        # pax(0, {"gbuf_views": [g for item in gbuf_view_items for g in item[2]]})
Lawrence McAfee's avatar
Lawrence McAfee committed
1383

1384
1385
1386
1387
1388
1389
1390
    def gather_params(self, ITERATION):

        # >>>
        timers = get_timers()
        # <<<

        timers('backward-params-all-gather').start()
Lawrence McAfee's avatar
Lawrence McAfee committed
1391

1392
1393
        data_parallel_rank = mpu.get_data_parallel_rank()
        data_parallel_group = mpu.get_data_parallel_group()
1394

1395
1396
        gbuf_view_items = self.get_model_grad_buffer_dp_views()

Lawrence McAfee's avatar
Lawrence McAfee committed
1397
        # All-gather updated main params.
1398
1399
1400
1401
1402
1403
1404
        for model_index, dtype, gbuf_views in gbuf_view_items:
            torch.distributed.all_gather(
                gbuf_views,
                gbuf_views[data_parallel_rank],
                group = data_parallel_group,
            )

1405
        # Each model param now contains its updated values in its
Lawrence McAfee's avatar
Lawrence McAfee committed
1406
        # '.main_grad' field.
1407
1408
1409
1410
1411
1412
1413
        # for param in self.param_gbuf_map: # ... incomplete param list.
        for model in self.models:
            for dtype, param_map in model._grad_buffer_param_index_map.items():
                for param in param_map:
                    param.detach().copy_(param.main_grad)

        timers('backward-params-all-gather').stop()
1414

1415
        # pax(0, {"gbuf_view_items": gbuf_view_items})
1416

1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
        # >>>
        # self.debug_main(ITERATION, "after/inside gather_params.", 0)
        # self.debug_model(ITERATION, "after/inside gather_params.", 0)

        # if ITERATION == 2:
        #     pax(1, {
        #         "ITERATION" : ITERATION,
        #         # "gbufs" : [
        #         #     tp(b.data)
        #         #     for m in self.models
        #         #     for b in m._grad_buffers.values()
        #         # ],
        #         "param_gbuf_map" : [ str(tuple(p.shape)) for p in self.param_gbuf_map ],
        #     })
        # <<<

Lawrence McAfee's avatar
Lawrence McAfee committed
1433
    def _collect_main_grad_data_for_unscaling(self):
1434
        return [ g.data for g in self.get_main_grads() ]
Lawrence McAfee's avatar
Lawrence McAfee committed
1435

1436
1437
1438
    def _copy_model_params_to_main_params(self):

        for group_index, group_shard in enumerate(self.opt_group_shards):
1439
            main_param = self.get_main_param(group_index)
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
            for model_param, main_shard in group_shard["param_map"].items():

                # Model shard.
                model_index, dtype = self.param_gbuf_map[model_param]
                model_shard = self.model_gbuf_shards \
                    [model_index][dtype]["param_map"][model_param]["param"]

                assert main_shard.size == model_shard.size

                # Copy shard data.
                main_view = main_param[main_shard.start:main_shard.end]
1451
                model_view = model_param.view(-1)[model_shard.start:model_shard.end]
1452

1453
                main_view.detach().copy_(model_view)
1454

1455

1456
1457
    def _copy_model_grads_to_main_grads(self, ITERATION):

Lawrence McAfee's avatar
Lawrence McAfee committed
1458
        for group_index, group_shard in enumerate(self.opt_group_shards):
1459
            for model_param, main_shard in group_shard["param_map"].items():
Lawrence McAfee's avatar
Lawrence McAfee committed
1460

1461
                # Model shard.
1462
                model_index, dtype = self.param_gbuf_map[model_param]
Lawrence McAfee's avatar
Lawrence McAfee committed
1463
                model_shard = self.model_gbuf_shards \
1464
                    [model_index][dtype]["param_map"][model_param]["gbuf_world"]
Lawrence McAfee's avatar
Lawrence McAfee committed
1465
1466
1467

                assert main_shard.size == model_shard.size

1468
1469
1470
1471
1472
1473
1474
                # pax(0, {
                #     "model_param" : tp(model_param),
                #     "main_shard" : str(main_shard),
                #     "param shard" : self.model_gbuf_shards \
                #     [model_index][dtype]["param_map"][model_param],
                # })

Lawrence McAfee's avatar
Lawrence McAfee committed
1475
                # Copy from DDP's contiguous buffer to main shard's grad.
1476
                model_grad = self.models[model_index]._grad_buffers[dtype].data
1477
                main_grad = self.get_main_grad(group_index)
Lawrence McAfee's avatar
Lawrence McAfee committed
1478

Lawrence McAfee's avatar
Lawrence McAfee committed
1479
                # Copy sub-range within tensor.
1480
1481
                model_view = model_grad[model_shard.start:model_shard.end]
                main_view = main_grad[main_shard.start:main_shard.end]
Lawrence McAfee's avatar
Lawrence McAfee committed
1482

1483
                main_view.detach().copy_(model_view)
Lawrence McAfee's avatar
Lawrence McAfee committed
1484
1485
1486
1487

                # pax(0, {
                #     "group_index" : group_index,
                #     "group_shard" : group_shard,
1488
                #     # "param" : tp(param),
Lawrence McAfee's avatar
Lawrence McAfee committed
1489
                #     "model_index" : model_index,
1490
1491
1492
1493
1494
                #     "dtype" : str(dtype),
                #     "model_grad" : tp(model_grad),
                #     "main_grad" : tp(main_grad),
                #     "model_view" : tp(model_view),
                #     "main_view" : tp(main_view),
Lawrence McAfee's avatar
Lawrence McAfee committed
1495
1496
1497
1498
                #     "model_shard" : str(model_shard),
                #     "main_shard" : str(main_shard),
                # })

Lawrence McAfee's avatar
Lawrence McAfee committed
1499
        # >>>
1500
        # if 1 or ITERATION == DEBUG_ITERATION:
1501
1502
1503
1504
1505
        #     pax(0, {
        #         "** branch **" : "** fix. **",
        #         "ITERATION" : ITERATION,
        #         # "model grads" : self.get_world_model_grads(),
        #         "main_grads" : self.get_main_grads(),
1506
1507
1508
1509
1510
        #         "group shards" : [
        #             "group %d; %s" % (grp_idx, main_shard)
        #             for grp_idx, grp_shard in enumerate(self.opt_group_shards)
        #             for model_param, main_shard in grp_shard["param_map"].items()
        #         ],
1511
        #     })
Lawrence McAfee's avatar
Lawrence McAfee committed
1512
        # <<<
Lawrence McAfee's avatar
Lawrence McAfee committed
1513

1514

1515
    def _copy_main_params_to_model_params(self, ITERATION):
1516
1517

        for group_index, group_shard in enumerate(self.opt_group_shards):
1518
            for model_param, main_shard in group_shard["param_map"].items():
1519

1520
                model_index, dtype = self.param_gbuf_map[model_param]
1521
                model_shard = self.model_gbuf_shards \
1522
                    [model_index][dtype]["param_map"][model_param]["gbuf_world"]
1523
1524
1525
1526

                assert main_shard.size == model_shard.size

                # Use DDP's contiguous buffer to temporarily hold params.
1527
                model_param = self.models[model_index]._grad_buffers[dtype].data
1528
                main_param = self.get_main_param(group_index)
1529
1530

                # Copy sub-range within tensor.
1531
1532
                model_view = model_param[model_shard.start:model_shard.end]
                main_view = main_param[main_shard.start:main_shard.end]
1533
1534
1535
1536

                model_view.detach().copy_(main_view)

                # Debug.
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
                # pax(1, {
                #     "group_index" : group_index,
                #     "group_shard" : group_shard,
                #     "model_param" : tp(model_param),
                #     "model_index" : model_index,
                #     "dtype" : str(dtype),
                #     "model_param" : tp(model_param),
                #     "main_param" : tp(main_param),
                #     "model_view" : tp(model_view),
                #     "main_view" : tp(main_view),
                #     "model_shard" : str(model_shard),
                #     "main_shard" : str(main_shard),
                # })
1550

Lawrence McAfee's avatar
Lawrence McAfee committed
1551
        # >>>
Lawrence McAfee's avatar
Lawrence McAfee committed
1552
1553
1554
1555
1556
1557
        # if ITERATION == DEBUG_ITERATION:
        #     pax(0, {
        #         "** branch **" : "** fix. **",
        #         "ITERATION" : ITERATION,
        #         "model params" : self.get_world_model_params(),
        #     })
Lawrence McAfee's avatar
Lawrence McAfee committed
1558
        # <<<
1559

1560
1561
# <<<

mohammad's avatar
mohammad committed
1562

mohammad's avatar
mohammad committed
1563
1564
class FP32Optimizer(MegatronOptimizer):

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
1565
1566
    def __init__(self, optimizer, clip_grad,
                 log_num_zeros_in_grad,
1567
                 params_have_main_grad,
1568
                 use_contiguous_buffers_in_local_ddp):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
1569
1570
1571

        super(FP32Optimizer, self).__init__(
            optimizer, clip_grad, log_num_zeros_in_grad,
1572
            params_have_main_grad, use_contiguous_buffers_in_local_ddp)
mohammad's avatar
mohammad committed
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590

        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
1591
        Always return successful since there is no overflow."""
mohammad's avatar
mohammad committed
1592

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
1593
1594
1595
1596
1597
1598
        # 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

1599
1600
1601
                    # 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.)
1602
                    if not self.use_contiguous_buffers_in_local_ddp:
1603
1604
                        param.main_grad = None

mohammad's avatar
mohammad committed
1605
        # Clip gradients.
1606
        grad_norm = None
mohammad's avatar
mohammad committed
1607
        if self.clip_grad > 0.0:
1608
            grad_norm = self.clip_grad_norm(self.clip_grad)
mohammad's avatar
mohammad committed
1609

Rewon Child's avatar
Rewon Child committed
1610
        # count the zeros in the grads
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
1611
1612
        num_zeros_in_grad = self.count_zeros() if \
                            self.log_num_zeros_in_grad else None
Rewon Child's avatar
Rewon Child committed
1613

mohammad's avatar
mohammad committed
1614
1615
1616
1617
        # Update parameters.
        self.optimizer.step()

        # No overflow for FP32 optimizer.
1618
        return True, grad_norm, num_zeros_in_grad
mohammad's avatar
mohammad committed
1619
1620


1621
1622
1623
1624
    def reload_model_params(self):
        pass


mohammad's avatar
mohammad committed
1625
1626
1627
1628
1629
1630
    def state_dict(self):
        return self.optimizer.state_dict()


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