File size: 13,731 Bytes
62e9ca6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# --------------------------------------------------------
# Copyright (c) 2022 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Based on fairseq code bases
# https://github.com/facebookresearch/fairseq
# --------------------------------------------------------

from typing import Dict, Optional, Tuple

import torch
import torch.nn.functional as F
from fairseq import utils
from torch import Tensor

from fairseq.modules import MultiheadAttention as FairseqMultiheadAttention


class MultiheadAttention(FairseqMultiheadAttention):
    """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,
        scaling_for_att=1.0
    ):
        super().__init__(
            embed_dim,
            num_heads,
            kdim,
            vdim,
            dropout,
            bias,
            add_bias_kv,
            add_zero_attn,
            self_attention,
            encoder_decoder_attention,
            q_noise,
            qn_block_size,
        )
        self.scaling_for_att = scaling_for_att

    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]]:
        """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 position_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:
            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
        q *= (1 / self.scaling_for_att)

        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)

        if position_bias is not None: ## first order
            ## position_bias: [241, 241, 64]
            #print ("attn_weights: ", attn_weights.size()) # [492, 241, 241]
            reshape_q = q.contiguous().view(bsz * self.num_heads, -1, self.head_dim).transpose(0,1) #[241, 492, 64]
            #print ("reshape_q: ", reshape_q.size())
            B = torch.matmul(reshape_q, position_bias.transpose(-2, -1))
            #print ("B: ", B.size())  ## [241, 492, 241]
            #B = B.transpose(0, 1).view(bsz, self.num_heads, position_bias.size(0), position_bias.size(1))
            B = B.transpose(0, 1).view(bsz*self.num_heads, position_bias.size(0), position_bias.size(1))
            #print ("B 2: ", B.size())
            attn_weights += B

        attn_weights *= self.scaling_for_att
        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)
            if self.onnx_trace:
                attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1)
            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 self.scaling_for_att > 1.0:
            attn_weights = attn_weights - attn_weights.detach().max(dim=-1, keepdim=True)[0]
        
        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)
        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