File size: 20,854 Bytes
ce922b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import math
from typing import Dict, Optional, Tuple

import torch
import torch.nn.functional as F
from fairseq import utils
from fairseq.incremental_decoding_utils import with_incremental_state
from fairseq.modules.fairseq_dropout import FairseqDropout
from fairseq.modules.quant_noise import quant_noise
from torch import Tensor, nn
from torch.nn import Parameter


@with_incremental_state
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,
        scale_factor=2,
        scale_heads=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 = FairseqDropout(
            dropout, module_name=self.__class__.__name__
        )

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

        self.self_attention = self_attention
        self.encoder_decoder_attention = encoder_decoder_attention
        self.c_attn = nn.Parameter(torch.ones((self.num_heads,)), requires_grad=True) if scale_heads else None

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

        self.k_proj = quant_noise(
            nn.Linear(self.kdim, embed_dim, bias=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, 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.reset_parameters()

        self.onnx_trace = False

    def prepare_for_onnx_export_(self):
        self.onnx_trace = True

    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)

    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,
        self_attn_mask: Optional[Tensor] = None,
        before_softmax: bool = False,
        need_head_weights: bool = False,
        attn_bias: Optional[Tensor] = None
    ) -> Tuple[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, f"query dim {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 (
            not self.onnx_trace
            and 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_attn_mask is None
            and attn_bias is None
        ):
            assert key is not None and value is not None
            return 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 or self.dropout_module.apply_during_inference,
                key_padding_mask,
                need_weights,
                attn_mask,
                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,
            )

        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 and self_attn_mask is None:
            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.head_dim)
            .transpose(0, 1)
        )
        if k is not None:
            k = (
                k.contiguous()
                .view(-1, bsz * self.num_heads, self.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_bias is not None:
            attn_weights += attn_bias

        if attn_mask is not None:
            attn_mask = attn_mask.unsqueeze(0)
            if self.onnx_trace:
                attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1)
            attn_weights += attn_mask

        if self_attn_mask is not None:
            self_attn_mask = self_attn_mask.unsqueeze(1).expand(bsz, self.num_heads, tgt_len, src_len)
            attn_weights += self_attn_mask.contiguous().view(bsz * self.num_heads, tgt_len, src_len)

        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

        attn_weights_float = utils.softmax(
            attn_weights, dim=-1, onnx_trace=self.onnx_trace
        )
        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]
        if self.onnx_trace and attn.size(1) == 1:
            # when ONNX tracing a single decoder step (sequence length == 1)
            # the transpose is a no-op copy before view, thus unnecessary
            attn = attn.contiguous().view(tgt_len, bsz, embed_dim)
        else:
            attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
        if self.c_attn is not None:
            attn = attn.view(tgt_len, bsz, self.num_heads, self.head_dim)
            attn = torch.einsum('tbhd,h->tbhd', attn, self.c_attn)
            attn = attn.reshape(tgt_len, bsz, self.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

    @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

    @torch.jit.export
    def reorder_incremental_state(
        self,
        incremental_state: Dict[str, Dict[str, Optional[Tensor]]],
        new_order: Tensor,
    ):
        """Reorder buffered internal state (for incremental generation)."""
        input_buffer = self._get_input_buffer(incremental_state)
        if input_buffer is not None:
            for k in input_buffer.keys():
                input_buffer_k = input_buffer[k]
                if input_buffer_k is not None:
                    if self.encoder_decoder_attention and input_buffer_k.size(
                        0
                    ) == new_order.size(0):
                        break
                    input_buffer[k] = input_buffer_k.index_select(0, new_order)
            incremental_state = self._set_input_buffer(incremental_state, input_buffer)
        return incremental_state

    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

    def upgrade_state_dict_named(self, state_dict, name):
        prefix = name + "." if name != "" else ""
        items_to_add = {}
        keys_to_remove = []
        for k in state_dict.keys():
            if k.endswith(prefix + "in_proj_weight"):
                # in_proj_weight used to be q + k + v with same dimensions
                dim = int(state_dict[k].shape[0] / 3)
                items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim]
                items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim]
                items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :]

                keys_to_remove.append(k)

                k_bias = prefix + "in_proj_bias"
                if k_bias in state_dict.keys():
                    dim = int(state_dict[k].shape[0] / 3)
                    items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim]
                    items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][
                        dim : 2 * dim
                    ]
                    items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :]

                    keys_to_remove.append(prefix + "in_proj_bias")

        for k in keys_to_remove:
            del state_dict[k]

        for key, value in items_to_add.items():
            state_dict[key] = value