layers.py 21.9 KB
Newer Older
1
# coding=utf-8
Mohammad's avatar
Mohammad committed
2
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
#
# 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.


# Parts of the code here are adapted from PyTorch
# repo: https://github.com/pytorch/pytorch


import math

import torch
import torch.nn.functional as F
import torch.nn.init as init
from torch.nn.parameter import Parameter

28
29
from .initialize import get_tensor_model_parallel_rank
from .initialize import get_tensor_model_parallel_world_size
30
from .initialize import get_tensor_model_parallel_group
31
from .mappings import copy_to_tensor_model_parallel_region
32
33
from .mappings import gather_from_tensor_model_parallel_region
from .mappings import gather_from_sequence_parallel_region
34
from .mappings import reduce_from_tensor_model_parallel_region
35
36
from .mappings import scatter_to_tensor_model_parallel_region
from .mappings import reduce_scatter_to_sequence_parallel_region
37

38
39
40
41
from .random import get_cuda_rng_tracker
from .utils import divide
from .utils import split_tensor_along_last_dim
from .utils import VocabUtility
42
from megatron import get_args
43

mohammad's avatar
mohammad committed
44
45
46
47
48
49

_MODEL_PARALLEL_ATTRIBUTE_DEFAULTS = {'tensor_model_parallel': False,
                                      'partition_dim': -1,
                                      'partition_stride': 1}


mohammad's avatar
mohammad committed
50
51
52
53
54
55
def param_is_not_tensor_parallel_duplicate(param):
    return (hasattr(param, 'tensor_model_parallel') and
            param.tensor_model_parallel) or (
                get_tensor_model_parallel_rank() == 0)


mohammad's avatar
mohammad committed
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def set_tensor_model_parallel_attributes(tensor, is_parallel, dim, stride):
    # Make sure the attributes are not set.
    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
        assert not hasattr(tensor, attribute)
    # Set the attributes.
    setattr(tensor, 'tensor_model_parallel', is_parallel)
    setattr(tensor, 'partition_dim', dim)
    setattr(tensor, 'partition_stride', stride)


def set_defaults_if_not_set_tensor_model_parallel_attributes(tensor):
    def maybe_set(attribute, value):
        if not hasattr(tensor, attribute):
            setattr(tensor, attribute, value)
    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
        maybe_set(attribute, _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS[attribute])


def copy_tensor_model_parallel_attributes(destination_tensor, source_tensor):
    def maybe_copy(attribute):
        if hasattr(source_tensor, attribute):
            setattr(destination_tensor, attribute,
                    getattr(source_tensor, attribute))
    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
        maybe_copy(attribute)


83
84
85
86
def _initialize_affine_weight_gpu(weight, init_method,
                                  partition_dim, stride=1):
    """Initialize affine weight for model parallel on GPU."""

mohammad's avatar
mohammad committed
87
88
89
90
91
    set_tensor_model_parallel_attributes(tensor=weight,
                                         is_parallel=True,
                                         dim=partition_dim,
                                         stride=stride)

92
93
94
95
96
97
98
99
    with get_cuda_rng_tracker().fork():
        init_method(weight)


def _initialize_affine_weight_cpu(weight, output_size, input_size,
                                  per_partition_size, partition_dim,
                                  init_method, stride=1,
                                  return_master_weight=False):
100
101
102
103
    """Initialize affine weight for model parallel.

    Build the master weight on all processes and scatter
    the relevant chunk."""
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
104

mohammad's avatar
mohammad committed
105
106
107
108
    set_tensor_model_parallel_attributes(tensor=weight,
                                         is_parallel=True,
                                         dim=partition_dim,
                                         stride=stride)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
109

110
111
    # Initialize master weight
    master_weight = torch.empty(output_size, input_size,
112
                                dtype=torch.float,
113
114
                                requires_grad=False)
    init_method(master_weight)
115
116
    args = get_args()
    master_weight = master_weight.to(dtype=args.params_dtype)
117
118
119
120
121

    # Split and copy
    per_partition_per_stride_size = divide(per_partition_size, stride)
    weight_list = torch.split(master_weight, per_partition_per_stride_size,
                              dim=partition_dim)
Jared Casper's avatar
Jared Casper committed
122
    rank = get_tensor_model_parallel_rank()
123
    world_size = get_tensor_model_parallel_world_size()
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
    my_weight_list = weight_list[rank::world_size]

    with torch.no_grad():
        torch.cat(my_weight_list, dim=partition_dim, out=weight)
    if return_master_weight:
        return master_weight
    return None


class VocabParallelEmbedding(torch.nn.Module):
    """Embedding parallelized in the vocabulary dimension.

    This is mainly adapted from torch.nn.Embedding and all the default
    values are kept.
    Arguments:
        num_embeddings: vocabulary size.
        embedding_dim: size of hidden state.
        init_method: method to initialize weights.
    """
Neel Kant's avatar
Neel Kant committed
143

144
145
146
147
148
149
150
151
152
153
154
155
156
    def __init__(self, num_embeddings, embedding_dim,
                 init_method=init.xavier_normal_):
        super(VocabParallelEmbedding, self).__init__()
        # Keep the input dimensions.
        self.num_embeddings = num_embeddings
        self.embedding_dim = embedding_dim
        # Set the detauls for compatibility.
        self.padding_idx = None
        self.max_norm = None
        self.norm_type = 2.
        self.scale_grad_by_freq = False
        self.sparse = False
        self._weight = None
157
        self.tensor_model_parallel_size = get_tensor_model_parallel_world_size()
158
159
160
        # Divide the weight matrix along the vocaburaly dimension.
        self.vocab_start_index, self.vocab_end_index = \
            VocabUtility.vocab_range_from_global_vocab_size(
161
162
                self.num_embeddings, get_tensor_model_parallel_rank(),
                self.tensor_model_parallel_size)
163
        self.num_embeddings_per_partition = self.vocab_end_index - \
Neel Kant's avatar
Neel Kant committed
164
            self.vocab_start_index
165

166
167
        # Allocate weights and initialize.
        args = get_args()
168
        if args.use_cpu_initialization:
169
170
171
172
173
174
175
176
177
178
179
180
            self.weight = Parameter(torch.empty(
                self.num_embeddings_per_partition, self.embedding_dim,
                dtype=args.params_dtype))
            _initialize_affine_weight_cpu(
                self.weight, self.num_embeddings, self.embedding_dim,
                self.num_embeddings_per_partition, 0, init_method)
        else:
            self.weight = Parameter(torch.empty(
                self.num_embeddings_per_partition, self.embedding_dim,
                device=torch.cuda.current_device(), dtype=args.params_dtype))
            _initialize_affine_weight_gpu(self.weight, init_method,
                                          partition_dim=0, stride=1)
181
182

    def forward(self, input_):
183
        if self.tensor_model_parallel_size > 1:
184
185
186
187
188
189
190
191
192
            # Build the mask.
            input_mask = (input_ < self.vocab_start_index) | \
                         (input_ >= self.vocab_end_index)
            # Mask the input.
            masked_input = input_.clone() - self.vocab_start_index
            masked_input[input_mask] = 0
        else:
            masked_input = input_
            # Get the embeddings.
193
194
195
196
197
        output_parallel = F.embedding(masked_input, self.weight,
                                      self.padding_idx, self.max_norm,
                                      self.norm_type, self.scale_grad_by_freq,
                                      self.sparse)
        # Mask the output embedding.
198
        if self.tensor_model_parallel_size > 1:
199
            output_parallel[input_mask, :] = 0.0
200
        # Reduce across all the model parallel GPUs.
201
        output = reduce_from_tensor_model_parallel_region(output_parallel)
202
203
204
        return output


205
class LinearWithGradAccumulationAndAsyncCommunication(torch.autograd.Function):
206
    """
207
208
    Linear layer execution with asynchronous communication and gradient accumulation
    fusion in backprop.
209
210
    """
    @staticmethod
211
212
    def forward(ctx, input, weight, bias, gradient_accumulation_fusion,
                async_grad_allreduce, model_parallel_memory_opt):
213
        ctx.save_for_backward(input, weight)
slym's avatar
slym committed
214
        ctx.use_bias = bias is not None
215
216
217
218
219
220
221
222
223
224
225
226
227
228
        ctx.gradient_accumulation_fusion = gradient_accumulation_fusion
        ctx.async_grad_allreduce = async_grad_allreduce
        ctx.model_parallel_memory_opt = model_parallel_memory_opt
        
        if model_parallel_memory_opt:
            world_size = get_tensor_model_parallel_world_size()
            dim_size = list(input.size())
            dim_size[0] = dim_size[0] * world_size

            total_input = torch.empty(dim_size, dtype=input.dtype,
                                      device=torch.cuda.current_device(),
                                      requires_grad=False)
            torch.distributed._all_gather_base(total_input, input,
                                               group=get_tensor_model_parallel_group())
229
        
230
231
232
        else:
            total_input = input

233
234
235
236
237
238
239
240
241
242
        output = torch.matmul(total_input, weight.t())
        if bias is not None:
            output = output + bias
        return output

    @staticmethod
    def backward(ctx, grad_output):
        input, weight = ctx.saved_tensors
        use_bias = ctx.use_bias
        
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
        if ctx.model_parallel_memory_opt:
            world_size = get_tensor_model_parallel_world_size()
            dim_size = list(input.size())
            dim_size[0] = dim_size[0] * world_size

            total_input = torch.empty(dim_size, dtype=input.dtype,
                                      device=torch.cuda.current_device(),
                                      requires_grad=False)
            handle = torch.distributed._all_gather_base(total_input, input,
                                           group=get_tensor_model_parallel_group(), async_op=True)
            # Delay the start of intput gradient computation shortly (3us) to have
            # gather scheduled first and have GPU resources allocated
            _ = torch.empty(1, device=grad_output.device) + 1
        else:
            total_input = input
258
259
        grad_input = grad_output.matmul(weight)

260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
        if ctx.model_parallel_memory_opt:
            handle.wait()

        # Convert the tensor shapes to 2D for execution compatibility
        grad_output = grad_output.view(grad_output.shape[0] * grad_output.shape[1],
                                       grad_output.shape[2])
        total_input = total_input.view(total_input.shape[0] * total_input.shape[1],
				       total_input.shape[2])
 
        if ctx.async_grad_allreduce:
            # Asynchronous all-reduce
            handle = torch.distributed.all_reduce(
                    grad_input, group=get_tensor_model_parallel_group(), async_op=True)
            # Delay the start of weight gradient computation shortly (3us) to have
            # all-reduce scheduled first and have GPU resources allocated
            _ = torch.empty(1, device=grad_output.device) + 1
 
        if ctx.model_parallel_memory_opt:
            assert not ctx.async_grad_allreduce
            dim_size = list(input.size())
            sub_grad_input = torch.empty(dim_size, dtype=input.dtype,
281
282
283
                             device=torch.cuda.current_device(),
                             requires_grad=False)

284
285
286
287
            # reduce_scatter
            handle = torch.distributed._reduce_scatter_base(sub_grad_input, grad_input, 
                                                            group=get_tensor_model_parallel_group(),
                                                            async_op=True)
288

289
290
291
292
293
294
295
296
297
298
            # Delay the start of weight gradient computation shortly (3us) to have
            # reduce scatter scheduled first and have GPU resources allocated
            _ = torch.empty(1, device=grad_output.device) + 1
        

        if ctx.gradient_accumulation_fusion:
            fused_dense_cuda.wgrad_gemm_accum_fp32(total_input, grad_output, weight.main_grad)
            grad_weight = None
        else:
            grad_weight = grad_output.t().matmul(total_input)
299
300
        grad_bias = grad_output.sum(dim=0) if use_bias else None

301
302
303
        if ctx.async_grad_allreducei:
            handle.wait()
            return grad_input, grad_weight, grad_bias
304

305
306
307
        if ctx.model_parallel_memory_opt:
            handle.wait()
            return sub_grad_input, grad_weight, grad_bias
308

309

310
311
312
313
314
315
316
317
318
319
class ColumnParallelLinear(torch.nn.Module):
    """Linear layer with column parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its second dimension as A = [A_1, ..., A_p].

    Arguments:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias
vycezhong's avatar
vycezhong committed
320
        gather_output: If true, call all-gather on output and make Y avaiable
321
322
323
324
325
326
327
328
                       to all GPUs, otherwise, every GPU will have its output
                       which is Y_i = XA_i
        init_method: method to initialize weights. Note that bias is always set
                     to zero.
        stride: For the strided linear layers.
        keep_master_weight_for_test: This was added for testing and should be
                                     set to False. It returns the master weights
                                     used for initialization.
329
330
331
        skip_bias_add: This was added to enable performance optimations where bias
                       can be fused with other elementwise operations. we skip 
                       adding bias but instead return it.
332
    """
Neel Kant's avatar
Neel Kant committed
333

334
335
    def __init__(self, input_size, output_size, bias=True, gather_output=True,
                 init_method=init.xavier_normal_, stride=1,
336
337
                 keep_master_weight_for_test=False,
                 skip_bias_add=False):
338
339
340
341
342
343
344
        super(ColumnParallelLinear, self).__init__()

        # Keep input parameters
        self.input_size = input_size
        self.output_size = output_size
        self.gather_output = gather_output
        # Divide the weight matrix along the last dimension.
345
        world_size = get_tensor_model_parallel_world_size()
346
        self.output_size_per_partition = divide(output_size, world_size)
347
        self.skip_bias_add = skip_bias_add
348
349
350
351

        # Parameters.
        # Note: torch.nn.functional.linear performs XA^T + b and as a result
        # we allocate the transpose.
352
353
        # Initialize weight.
        args = get_args()
354
        if args.use_cpu_initialization:
355
356
357
358
359
360
361
362
363
364
365
366
367
            self.weight = Parameter(torch.empty(self.output_size_per_partition,
                                                self.input_size,
                                                dtype=args.params_dtype))
            self.master_weight = _initialize_affine_weight_cpu(
                self.weight, self.output_size, self.input_size,
                self.output_size_per_partition, 0, init_method,
                stride=stride, return_master_weight=keep_master_weight_for_test)
        else:
            self.weight = Parameter(torch.empty(
                self.output_size_per_partition, self.input_size,
                device=torch.cuda.current_device(), dtype=args.params_dtype))
            _initialize_affine_weight_gpu(self.weight, init_method,
                                          partition_dim=0, stride=stride)
hwijeen's avatar
hwijeen committed
368

369
        if bias:
370
            if args.use_cpu_initialization:
371
372
373
374
375
376
377
                self.bias = Parameter(torch.empty(
                    self.output_size_per_partition, dtype=args.params_dtype))
            else:
                self.bias = Parameter(torch.empty(
                    self.output_size_per_partition,
                    device=torch.cuda.current_device(),
                    dtype=args.params_dtype))
378
            set_tensor_model_parallel_attributes(self.bias, True, 0, stride)
379
380
381
382
383
            # Always initialize bias to zero.
            with torch.no_grad():
                self.bias.zero_()
        else:
            self.register_parameter('bias', None)
slym's avatar
slym committed
384
        self.async_tensor_model_parallel_allreduce = (
slym's avatar
slym committed
385
386
                not args.no_async_tensor_model_parallel_allreduce and
                world_size > 1)
387
388
389
390
391
        self.model_parallel_memory_opt = (
                args.model_parallel_memory_opt and
                world_size > 1)
        assert not self.async_tensor_model_parallel_allreduce or \
            not self.model_parallel_memory_opt
392

393
394

    def forward(self, input_):
395
        bias = self.bias if not self.skip_bias_add else None
396

397
398
399
        if self.async_tensor_model_parallel_allreduce or \
                self.model_parallel_memory_opt:
            input_parallel = input_
400
        else:
401
402
403
404
405
            input_parallel = copy_to_tensor_model_parallel_region(input_)
        # Matrix multiply.
        output_parallel = LinearWithGradAccumulationAndAsyncCommunication.apply(
            input_parallel, self.weight, bias, self.gradient_accumulation_fusion,
            self.async_tensor_model_parallel_allreduce, self.model_parallel_memory_opt)
406
407
        if self.gather_output:
            # All-gather across the partitions.
408
            assert not self.model_parallel_memory_opt
409
            output = gather_from_tensor_model_parallel_region(output_parallel)
410
        else:
hwijeen's avatar
hwijeen committed
411
            output = output_parallel
412
413
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440


class RowParallelLinear(torch.nn.Module):
    """Linear layer with row parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its first dimension and X along its second dimension as:
               -   -
              | A_1 |
              | .   |
          A = | .   |        X = [X_1, ..., X_p]
              | .   |
              | A_p |
               -   -
    Arguments:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias. Note that bias is not parallelized.
        input_is_parallel: If true, we assume that the input is already
                           split across the GPUs and we do not split
                           again.
        init_method: method to initialize weights. Note that bias is always set
                     to zero.
        stride: For the strided linear layers.
        keep_master_weight_for_test: This was added for testing and should be
                                     set to False. It returns the master weights
                                     used for initialization.
hwijeen's avatar
hwijeen committed
441
442
        skip_bias_add: This was added to enable performance optimization where bias
                       can be fused with other elementwise operations. We skip
443
                       adding bias but instead return it.
444
    """
Neel Kant's avatar
Neel Kant committed
445

446
447
448
    def __init__(self, input_size, output_size, bias=True,
                 input_is_parallel=False,
                 init_method=init.xavier_normal_, stride=1,
449
450
                 keep_master_weight_for_test=False,
                 skip_bias_add=False):
451
452
453
454
455
456
457
        super(RowParallelLinear, self).__init__()

        # Keep input parameters
        self.input_size = input_size
        self.output_size = output_size
        self.input_is_parallel = input_is_parallel
        # Divide the weight matrix along the last dimension.
458
        world_size = get_tensor_model_parallel_world_size()
459
        self.input_size_per_partition = divide(input_size, world_size)
460
        self.skip_bias_add = skip_bias_add
461
462
463
464

        # Parameters.
        # Note: torch.nn.functional.linear performs XA^T + b and as a result
        # we allocate the transpose.
465
466
        # Initialize weight.
        args = get_args()
467
        if args.use_cpu_initialization:
468
469
470
471
472
473
474
475
476
477
478
479
480
            self.weight = Parameter(torch.empty(self.output_size,
                                                self.input_size_per_partition,
                                                dtype=args.params_dtype))
            self.master_weight = _initialize_affine_weight_cpu(
                self.weight, self.output_size, self.input_size,
                self.input_size_per_partition, 1, init_method,
                stride=stride, return_master_weight=keep_master_weight_for_test)
        else:
            self.weight = Parameter(torch.empty(
                self.output_size, self.input_size_per_partition,
                device=torch.cuda.current_device(), dtype=args.params_dtype))
            _initialize_affine_weight_gpu(self.weight, init_method,
                                          partition_dim=1, stride=stride)
481
        if bias:
482
            if args.use_cpu_initialization:
483
484
485
486
487
488
                self.bias = Parameter(torch.empty(self.output_size,
                                                  dtype=args.params_dtype))
            else:
                self.bias = Parameter(torch.empty(
                    self.output_size, device=torch.cuda.current_device(),
                    dtype=args.params_dtype))
489
490
491
492
493
494
            # Always initialize bias to zero.
            with torch.no_grad():
                self.bias.zero_()
        else:
            self.register_parameter('bias', None)

495
        self.model_parallel_memory_opt = args.model_parallel_memory_opt
496

497
498
499
500
501
502

    def forward(self, input_):
        # Set up backprop all-reduce.
        if self.input_is_parallel:
            input_parallel = input_
        else:
503
            assert not self.model_parallel_memory_opt
504
            input_parallel = scatter_to_tensor_model_parallel_region(input_)
505
506
507
        # Matrix multiply.
        output_parallel = F.linear(input_parallel, self.weight)
        # All-reduce across all the partitions.
508
        if self.model_parallel_memory_opt:
509
            output_ = reduce_scatter_to_sequence_parallel_region(output_parallel)
510
511
        else:
            output_ = reduce_from_tensor_model_parallel_region(output_parallel)
512
513
514
        if not self.skip_bias_add:
            output = output_ + self.bias if self.bias is not None else output_
            output_bias = None
515
516
        else:
            output = output_
517
518
519
            output_bias = self.bias
        return output, output_bias