File size: 31,945 Bytes
fc5ed00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
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
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
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
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
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
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
# --------------------------------------------------------
# WavLM: Large-Scale Self-Supervised  Pre-training  for Full Stack Speech Processing (https://arxiv.org/abs/2110.13900.pdf)
# Github source: https://github.com/microsoft/unilm/tree/master/wavlm
# Copyright (c) 2021 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Based on fairseq code bases
# https://github.com/pytorch/fairseq
# --------------------------------------------------------

import math
import warnings
from typing import Dict, Optional, Tuple
import torch
from torch import Tensor, nn
from torch.nn import Parameter
import torch.nn.functional as F


class TransposeLast(nn.Module):
    def __init__(self, deconstruct_idx=None):
        super().__init__()
        self.deconstruct_idx = deconstruct_idx

    def forward(self, x):
        if self.deconstruct_idx is not None:
            x = x[self.deconstruct_idx]
        return x.transpose(-2, -1)


class Fp32LayerNorm(nn.LayerNorm):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def forward(self, input):
        output = F.layer_norm(
            input.float(),
            self.normalized_shape,
            self.weight.float() if self.weight is not None else None,
            self.bias.float() if self.bias is not None else None,
            self.eps,
        )
        return output.type_as(input)


class Fp32GroupNorm(nn.GroupNorm):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def forward(self, input):
        output = F.group_norm(
            input.float(),
            self.num_groups,
            self.weight.float() if self.weight is not None else None,
            self.bias.float() if self.bias is not None else None,
            self.eps,
        )
        return output.type_as(input)


class GradMultiply(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x, scale):
        ctx.scale = scale
        res = x.new(x)
        return res

    @staticmethod
    def backward(ctx, grad):
        return grad * ctx.scale, None


class SamePad(nn.Module):
    def __init__(self, kernel_size, causal=False):
        super().__init__()
        if causal:
            self.remove = kernel_size - 1
        else:
            self.remove = 1 if kernel_size % 2 == 0 else 0

    def forward(self, x):
        if self.remove > 0:
            x = x[:, :, : -self.remove]
        return x


class Swish(nn.Module):
    """Swish function
    """

    def __init__(self):
        """Construct an MultiHeadedAttention object."""
        super(Swish, self).__init__()
        self.act = torch.nn.Sigmoid()

    def forward(self, x):
        return x * self.act(x)


class GLU_Linear(nn.Module):
    def __init__(self, input_dim, output_dim, glu_type="sigmoid", bias_in_glu=True):
        super(GLU_Linear, self).__init__()

        self.glu_type = glu_type
        self.output_dim = output_dim

        if glu_type == "sigmoid":
            self.glu_act = torch.nn.Sigmoid()
        elif glu_type == "swish":
            self.glu_act = Swish()
        elif glu_type == "relu":
            self.glu_act = torch.nn.ReLU()
        elif glu_type == "gelu":
            self.glu_act = torch.nn.GELU()

        if bias_in_glu:
            self.linear = nn.Linear(input_dim, output_dim * 2, True)
        else:
            self.linear = nn.Linear(input_dim, output_dim * 2, False)

    def forward(self, x):
        # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case
        x = self.linear(x)

        if self.glu_type == "bilinear":
            x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])
        else:
            x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))

        return x


def gelu_accurate(x):
    if not hasattr(gelu_accurate, "_a"):
        gelu_accurate._a = math.sqrt(2 / math.pi)
    return (
        0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))
    )


def gelu(x: torch.Tensor) -> torch.Tensor:
    return torch.nn.functional.gelu(x.float()).type_as(x)


def get_activation_fn(activation: str):
    """Returns the activation function corresponding to `activation`"""

    if activation == "relu":
        return F.relu
    elif activation == "gelu":
        return gelu
    elif activation == "gelu_fast":
        warnings.warn(
            "--activation-fn=gelu_fast has been renamed to gelu_accurate"
        )
        return gelu_accurate
    elif activation == "gelu_accurate":
        return gelu_accurate
    elif activation == "tanh":
        return torch.tanh
    elif activation == "linear":
        return lambda x: x
    elif activation == "glu":
        return lambda x: x
    else:
        raise RuntimeError("--activation-fn {} not supported".format(activation))


def init_bert_params(module):
    """
    Initialize the weights specific to the BERT Model.
    This overrides the default initializations depending on the specified arguments.
        1. If normal_init_linear_weights is set then weights of linear
           layer will be initialized using the normal distribution and
           bais will be set to the specified value.
        2. If normal_init_embed_weights is set then weights of embedding
           layer will be initialized using the normal distribution.
        3. If normal_init_proj_weights is set then weights of
           in_project_weight for MultiHeadAttention initialized using
           the normal distribution (to be validated).
    """

    def normal_(data):
        # with FSDP, module params will be on CUDA, so we cast them back to CPU
        # so that the RNG is consistent with and without FSDP
        data.copy_(
            data.cpu().normal_(mean=0.0, std=0.02).to(data.device)
        )

    if isinstance(module, nn.Linear):
        normal_(module.weight.data)
        if module.bias is not None:
            module.bias.data.zero_()
    if isinstance(module, nn.Embedding):
        normal_(module.weight.data)
        if module.padding_idx is not None:
            module.weight.data[module.padding_idx].zero_()
    if isinstance(module, MultiheadAttention):
        normal_(module.q_proj.weight.data)
        normal_(module.k_proj.weight.data)
        normal_(module.v_proj.weight.data)


def quant_noise(module, p, block_size):
    """
    Wraps modules and applies quantization noise to the weights for
    subsequent quantization with Iterative Product Quantization as
    described in "Training with Quantization Noise for Extreme Model Compression"

    Args:
        - module: nn.Module
        - p: amount of Quantization Noise
        - block_size: size of the blocks for subsequent quantization with iPQ

    Remarks:
        - Module weights must have the right sizes wrt the block size
        - Only Linear, Embedding and Conv2d modules are supported for the moment
        - For more detail on how to quantize by blocks with convolutional weights,
          see "And the Bit Goes Down: Revisiting the Quantization of Neural Networks"
        - We implement the simplest form of noise here as stated in the paper
          which consists in randomly dropping blocks
    """

    # if no quantization noise, don't register hook
    if p <= 0:
        return module

    # supported modules
    assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d))

    # test whether module.weight has the right sizes wrt block_size
    is_conv = module.weight.ndim == 4

    # 2D matrix
    if not is_conv:
        assert (
            module.weight.size(1) % block_size == 0
        ), "Input features must be a multiple of block sizes"

    # 4D matrix
    else:
        # 1x1 convolutions
        if module.kernel_size == (1, 1):
            assert (
                module.in_channels % block_size == 0
            ), "Input channels must be a multiple of block sizes"
        # regular convolutions
        else:
            k = module.kernel_size[0] * module.kernel_size[1]
            assert k % block_size == 0, "Kernel size must be a multiple of block size"

    def _forward_pre_hook(mod, input):
        # no noise for evaluation
        if mod.training:
            if not is_conv:
                # gather weight and sizes
                weight = mod.weight
                in_features = weight.size(1)
                out_features = weight.size(0)

                # split weight matrix into blocks and randomly drop selected blocks
                mask = torch.zeros(
                    in_features // block_size * out_features, device=weight.device
                )
                mask.bernoulli_(p)
                mask = mask.repeat_interleave(block_size, -1).view(-1, in_features)

            else:
                # gather weight and sizes
                weight = mod.weight
                in_channels = mod.in_channels
                out_channels = mod.out_channels

                # split weight matrix into blocks and randomly drop selected blocks
                if mod.kernel_size == (1, 1):
                    mask = torch.zeros(
                        int(in_channels // block_size * out_channels),
                        device=weight.device,
                    )
                    mask.bernoulli_(p)
                    mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels)
                else:
                    mask = torch.zeros(
                        weight.size(0), weight.size(1), device=weight.device
                    )
                    mask.bernoulli_(p)
                    mask = (
                        mask.unsqueeze(2)
                        .unsqueeze(3)
                        .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1])
                    )

            # scale weights and apply mask
            mask = mask.to(
                torch.bool
            )  # x.bool() is not currently supported in TorchScript
            s = 1 / (1 - p)
            mod.weight.data = s * weight.masked_fill(mask, 0)

    module.register_forward_pre_hook(_forward_pre_hook)
    return module


class MultiheadAttention(nn.Module):
    """Multi-headed attention.

    See "Attention Is All You Need" for more details.
    """

    def __init__(
            self,
            embed_dim,
            num_heads,
            kdim=None,
            vdim=None,
            dropout=0.0,
            bias=True,
            add_bias_kv=False,
            add_zero_attn=False,
            self_attention=False,
            encoder_decoder_attention=False,
            q_noise=0.0,
            qn_block_size=8,
            has_relative_attention_bias=False,
            num_buckets=32,
            max_distance=128,
            gru_rel_pos=False,
            rescale_init=False,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.kdim = kdim if kdim is not None else embed_dim
        self.vdim = vdim if vdim is not None else embed_dim
        self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim

        self.num_heads = num_heads
        self.dropout_module = nn.Dropout(dropout)

        self.has_relative_attention_bias = has_relative_attention_bias
        self.num_buckets = num_buckets
        self.max_distance = max_distance
        if self.has_relative_attention_bias:
            self.relative_attention_bias = nn.Embedding(num_buckets, num_heads)

        self.head_dim = embed_dim // num_heads
        self.q_head_dim = self.head_dim
        self.k_head_dim = self.head_dim
        assert (
                self.head_dim * num_heads == self.embed_dim
        ), "embed_dim must be divisible by num_heads"
        self.scaling = self.head_dim ** -0.5

        self.self_attention = self_attention
        self.encoder_decoder_attention = encoder_decoder_attention

        assert not self.self_attention or self.qkv_same_dim, (
            "Self-attention requires query, key and " "value to be of the same size"
        )

        k_bias = True
        if rescale_init:
            k_bias = False

        k_embed_dim = embed_dim
        q_embed_dim = embed_dim

        self.k_proj = quant_noise(
            nn.Linear(self.kdim, k_embed_dim, bias=k_bias), q_noise, qn_block_size
        )
        self.v_proj = quant_noise(
            nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size
        )
        self.q_proj = quant_noise(
            nn.Linear(embed_dim, q_embed_dim, bias=bias), q_noise, qn_block_size
        )

        self.out_proj = quant_noise(
            nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size
        )

        if add_bias_kv:
            self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))
            self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))
        else:
            self.bias_k = self.bias_v = None

        self.add_zero_attn = add_zero_attn

        self.gru_rel_pos = gru_rel_pos
        if self.gru_rel_pos:
            self.grep_linear = nn.Linear(self.q_head_dim, 8)
            self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))

        self.reset_parameters()

    def reset_parameters(self):
        if self.qkv_same_dim:
            # Empirically observed the convergence to be much better with
            # the scaled initialization
            nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
            nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
            nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
        else:
            nn.init.xavier_uniform_(self.k_proj.weight)
            nn.init.xavier_uniform_(self.v_proj.weight)
            nn.init.xavier_uniform_(self.q_proj.weight)

        nn.init.xavier_uniform_(self.out_proj.weight)
        if self.out_proj.bias is not None:
            nn.init.constant_(self.out_proj.bias, 0.0)
        if self.bias_k is not None:
            nn.init.xavier_normal_(self.bias_k)
        if self.bias_v is not None:
            nn.init.xavier_normal_(self.bias_v)
        if self.has_relative_attention_bias:
            nn.init.xavier_normal_(self.relative_attention_bias.weight)

    def _relative_positions_bucket(self, relative_positions, bidirectional=True):
        num_buckets = self.num_buckets
        max_distance = self.max_distance
        relative_buckets = 0

        if bidirectional:
            num_buckets = num_buckets // 2
            relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets
            relative_positions = torch.abs(relative_positions)
        else:
            relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))

        max_exact = num_buckets // 2
        is_small = relative_positions < max_exact

        relative_postion_if_large = max_exact + (
                torch.log(relative_positions.float() / max_exact)
                / math.log(max_distance / max_exact)
                * (num_buckets - max_exact)
        ).to(torch.long)
        relative_postion_if_large = torch.min(
            relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)
        )

        relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)
        return relative_buckets

    def compute_bias(self, query_length, key_length):
        context_position = torch.arange(query_length, dtype=torch.long)[:, None]
        memory_position = torch.arange(key_length, dtype=torch.long)[None, :]
        relative_position = memory_position - context_position
        relative_position_bucket = self._relative_positions_bucket(
            relative_position,
            bidirectional=True
        )
        relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)
        values = self.relative_attention_bias(relative_position_bucket)
        values = values.permute([2, 0, 1])
        return values

    def forward(
            self,
            query,
            key: Optional[Tensor],
            value: Optional[Tensor],
            key_padding_mask: Optional[Tensor] = None,
            incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
            need_weights: bool = True,
            static_kv: bool = False,
            attn_mask: Optional[Tensor] = None,
            before_softmax: bool = False,
            need_head_weights: bool = False,
            position_bias: Optional[Tensor] = None
    ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:
        """Input shape: Time x Batch x Channel

        Args:
            key_padding_mask (ByteTensor, optional): mask to exclude
                keys that are pads, of shape `(batch, src_len)`, where
                padding elements are indicated by 1s.
            need_weights (bool, optional): return the attention weights,
                averaged over heads (default: False).
            attn_mask (ByteTensor, optional): typically used to
                implement causal attention, where the mask prevents the
                attention from looking forward in time (default: None).
            before_softmax (bool, optional): return the raw attention
                weights and values before the attention softmax.
            need_head_weights (bool, optional): return the attention
                weights for each head. Implies *need_weights*. Default:
                return the average attention weights over all heads.
        """
        if need_head_weights:
            need_weights = True

        is_tpu = query.device.type == "xla"

        tgt_len, bsz, embed_dim = query.size()
        src_len = tgt_len
        assert embed_dim == self.embed_dim
        assert list(query.size()) == [tgt_len, bsz, embed_dim]
        if key is not None:
            src_len, key_bsz, _ = key.size()
            if not torch.jit.is_scripting():
                assert key_bsz == bsz
                assert value is not None
                assert src_len, bsz == value.shape[:2]

        if self.has_relative_attention_bias and position_bias is None:
            position_bias = self.compute_bias(tgt_len, src_len)
            position_bias = position_bias.unsqueeze(0).repeat(bsz, 1, 1, 1).view(bsz * self.num_heads, tgt_len, src_len)

        if (
                not is_tpu  # don't use PyTorch version on TPUs
                and incremental_state is None
                and not static_kv
                # A workaround for quantization to work. Otherwise JIT compilation
                # treats bias in linear module as method.
                and not torch.jit.is_scripting()
                and self.q_head_dim == self.head_dim
        ):
            assert key is not None and value is not None
            assert attn_mask is None

            attn_mask_rel_pos = None
            if position_bias is not None:
                attn_mask_rel_pos = position_bias
                if self.gru_rel_pos:
                    query_layer = query.transpose(0, 1)
                    new_x_shape = query_layer.size()[:-1] + (self.num_heads, -1)
                    query_layer = query_layer.view(*new_x_shape)
                    query_layer = query_layer.permute(0, 2, 1, 3)
                    _B, _H, _L, __ = query_layer.size()

                    gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(
                        _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)
                    gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0
                    attn_mask_rel_pos = gate_a_1.view(bsz * self.num_heads, -1, 1) * position_bias

                attn_mask_rel_pos = attn_mask_rel_pos.view((-1, tgt_len, tgt_len))
            k_proj_bias = self.k_proj.bias
            if k_proj_bias is None:
                k_proj_bias = torch.zeros_like(self.q_proj.bias)

            x, attn = F.multi_head_attention_forward(
                query,
                key,
                value,
                self.embed_dim,
                self.num_heads,
                torch.empty([0]),
                torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)),
                self.bias_k,
                self.bias_v,
                self.add_zero_attn,
                self.dropout_module.p,
                self.out_proj.weight,
                self.out_proj.bias,
                self.training,
                # self.training or self.dropout_module.apply_during_inference,
                key_padding_mask,
                need_weights,
                attn_mask_rel_pos,
                use_separate_proj_weight=True,
                q_proj_weight=self.q_proj.weight,
                k_proj_weight=self.k_proj.weight,
                v_proj_weight=self.v_proj.weight,
            )
            return x, attn, position_bias

        if incremental_state is not None:
            saved_state = self._get_input_buffer(incremental_state)
            if saved_state is not None and "prev_key" in saved_state:
                # previous time steps are cached - no need to recompute
                # key and value if they are static
                if static_kv:
                    assert self.encoder_decoder_attention and not self.self_attention
                    key = value = None
        else:
            saved_state = None

        if self.self_attention:
            q = self.q_proj(query)
            k = self.k_proj(query)
            v = self.v_proj(query)
        elif self.encoder_decoder_attention:
            # encoder-decoder attention
            q = self.q_proj(query)
            if key is None:
                assert value is None
                k = v = None
            else:
                k = self.k_proj(key)
                v = self.v_proj(key)

        else:
            assert key is not None and value is not None
            q = self.q_proj(query)
            k = self.k_proj(key)
            v = self.v_proj(value)
        q *= self.scaling

        if self.bias_k is not None:
            assert self.bias_v is not None
            k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
            v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
            if attn_mask is not None:
                attn_mask = torch.cat(
                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
                )
            if key_padding_mask is not None:
                key_padding_mask = torch.cat(
                    [
                        key_padding_mask,
                        key_padding_mask.new_zeros(key_padding_mask.size(0), 1),
                    ],
                    dim=1,
                )

        q = (
            q.contiguous()
                .view(tgt_len, bsz * self.num_heads, self.q_head_dim)
                .transpose(0, 1)
        )
        if k is not None:
            k = (
                k.contiguous()
                    .view(-1, bsz * self.num_heads, self.k_head_dim)
                    .transpose(0, 1)
            )
        if v is not None:
            v = (
                v.contiguous()
                    .view(-1, bsz * self.num_heads, self.head_dim)
                    .transpose(0, 1)
            )

        if saved_state is not None:
            # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
            if "prev_key" in saved_state:
                _prev_key = saved_state["prev_key"]
                assert _prev_key is not None
                prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)
                if static_kv:
                    k = prev_key
                else:
                    assert k is not None
                    k = torch.cat([prev_key, k], dim=1)
                src_len = k.size(1)
            if "prev_value" in saved_state:
                _prev_value = saved_state["prev_value"]
                assert _prev_value is not None
                prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)
                if static_kv:
                    v = prev_value
                else:
                    assert v is not None
                    v = torch.cat([prev_value, v], dim=1)
            prev_key_padding_mask: Optional[Tensor] = None
            if "prev_key_padding_mask" in saved_state:
                prev_key_padding_mask = saved_state["prev_key_padding_mask"]
            assert k is not None and v is not None
            key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(
                key_padding_mask=key_padding_mask,
                prev_key_padding_mask=prev_key_padding_mask,
                batch_size=bsz,
                src_len=k.size(1),
                static_kv=static_kv,
            )

            saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim)
            saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim)
            saved_state["prev_key_padding_mask"] = key_padding_mask
            # In this branch incremental_state is never None
            assert incremental_state is not None
            incremental_state = self._set_input_buffer(incremental_state, saved_state)
        assert k is not None
        assert k.size(1) == src_len

        # This is part of a workaround to get around fork/join parallelism
        # not supporting Optional types.
        if key_padding_mask is not None and key_padding_mask.dim() == 0:
            key_padding_mask = None

        if key_padding_mask is not None:
            assert key_padding_mask.size(0) == bsz
            assert key_padding_mask.size(1) == src_len

        if self.add_zero_attn:
            assert v is not None
            src_len += 1
            k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)
            v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)
            if attn_mask is not None:
                attn_mask = torch.cat(
                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
                )
            if key_padding_mask is not None:
                key_padding_mask = torch.cat(
                    [
                        key_padding_mask,
                        torch.zeros(key_padding_mask.size(0), 1).type_as(
                            key_padding_mask
                        ),
                    ],
                    dim=1,
                )

        attn_weights = torch.bmm(q, k.transpose(1, 2))
        attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)

        assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]

        if attn_mask is not None:
            attn_mask = attn_mask.unsqueeze(0)
            attn_weights += attn_mask

        if key_padding_mask is not None:
            # don't attend to padding symbols
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            if not is_tpu:
                attn_weights = attn_weights.masked_fill(
                    key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
                    float("-inf"),
                )
            else:
                attn_weights = attn_weights.transpose(0, 2)
                attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf"))
                attn_weights = attn_weights.transpose(0, 2)
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        if before_softmax:
            return attn_weights, v, position_bias

        if position_bias is not None:
            if self.gru_rel_pos == 1:
                query_layer = q.view(bsz, self.num_heads, tgt_len, self.q_head_dim)
                _B, _H, _L, __ = query_layer.size()
                gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(
                    _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)
                gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0
                position_bias = gate_a_1.view(bsz * self.num_heads, -1, 1) * position_bias

            position_bias = position_bias.view(attn_weights.size())

            attn_weights = attn_weights + position_bias

        attn_weights_float = F.softmax(
            attn_weights, dim=-1
        )
        attn_weights = attn_weights_float.type_as(attn_weights)
        attn_probs = self.dropout_module(attn_weights)

        assert v is not None
        attn = torch.bmm(attn_probs, v)
        assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
        attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
        attn = self.out_proj(attn)
        attn_weights: Optional[Tensor] = None
        if need_weights:
            attn_weights = attn_weights_float.view(
                bsz, self.num_heads, tgt_len, src_len
            ).transpose(1, 0)
            if not need_head_weights:
                # average attention weights over heads
                attn_weights = attn_weights.mean(dim=0)

        return attn, attn_weights, position_bias

    @staticmethod
    def _append_prev_key_padding_mask(
            key_padding_mask: Optional[Tensor],
            prev_key_padding_mask: Optional[Tensor],
            batch_size: int,
            src_len: int,
            static_kv: bool,
    ) -> Optional[Tensor]:
        # saved key padding masks have shape (bsz, seq_len)
        if prev_key_padding_mask is not None and static_kv:
            new_key_padding_mask = prev_key_padding_mask
        elif prev_key_padding_mask is not None and key_padding_mask is not None:
            new_key_padding_mask = torch.cat(
                [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1
            )
        # During incremental decoding, as the padding token enters and
        # leaves the frame, there will be a time when prev or current
        # is None
        elif prev_key_padding_mask is not None:
            if src_len > prev_key_padding_mask.size(1):
                filler = torch.zeros(
                    (batch_size, src_len - prev_key_padding_mask.size(1)),
                    device=prev_key_padding_mask.device,
                )
                new_key_padding_mask = torch.cat(
                    [prev_key_padding_mask.float(), filler.float()], dim=1
                )
            else:
                new_key_padding_mask = prev_key_padding_mask.float()
        elif key_padding_mask is not None:
            if src_len > key_padding_mask.size(1):
                filler = torch.zeros(
                    (batch_size, src_len - key_padding_mask.size(1)),
                    device=key_padding_mask.device,
                )
                new_key_padding_mask = torch.cat(
                    [filler.float(), key_padding_mask.float()], dim=1
                )
            else:
                new_key_padding_mask = key_padding_mask.float()
        else:
            new_key_padding_mask = prev_key_padding_mask
        return new_key_padding_mask

    def _get_input_buffer(
            self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]
    ) -> Dict[str, Optional[Tensor]]:
        result = self.get_incremental_state(incremental_state, "attn_state")
        if result is not None:
            return result
        else:
            empty_result: Dict[str, Optional[Tensor]] = {}
            return empty_result

    def _set_input_buffer(
            self,
            incremental_state: Dict[str, Dict[str, Optional[Tensor]]],
            buffer: Dict[str, Optional[Tensor]],
    ):
        return self.set_incremental_state(incremental_state, "attn_state", buffer)

    def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):
        return attn_weights