File size: 23,953 Bytes
c705408
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init
import logging
from diffusers.models.attention import Attention
from diffusers.utils import USE_PEFT_BACKEND, is_xformers_available
from typing import Optional, Callable

from einops import rearrange

if is_xformers_available():
    import xformers
    import xformers.ops
else:
    xformers = None

logger = logging.getLogger(__name__)


class AttnProcessor:
    r"""
    Default processor for performing attention-related computations.
    """

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        temb: Optional[torch.FloatTensor] = None,
        scale: float = 1.0,
        pose_feature=None,  # the only difference to the original code
    ) -> torch.Tensor:
        residual = hidden_states

        args = () if USE_PEFT_BACKEND else (scale,)

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )
        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states, *args)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states, *args)
        value = attn.to_v(encoder_hidden_states, *args)

        query = attn.head_to_batch_dim(query)
        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        attention_probs = attn.get_attention_scores(query, key, attention_mask)
        hidden_states = torch.bmm(attention_probs, value)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states, *args)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class AttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        temb: Optional[torch.FloatTensor] = None,
        scale: float = 1.0,
        pose_feature=None
    ) -> torch.FloatTensor:
        residual = hidden_states

        args = () if USE_PEFT_BACKEND else (scale,)

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        args = () if USE_PEFT_BACKEND else (scale,)
        query = attn.to_q(hidden_states, *args)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states, *args)
        value = attn.to_v(encoder_hidden_states, *args)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states, *args)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class XFormersAttnProcessor:
    r"""
    Processor for implementing memory efficient attention using xFormers.

    Args:
        attention_op (`Callable`, *optional*, defaults to `None`):
            The base
            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
            operator.
    """

    def __init__(self, attention_op: Optional[Callable] = None):
        self.attention_op = attention_op

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        temb: Optional[torch.FloatTensor] = None,
        scale: float = 1.0,
        pose_feature=None,  # the only difference to the original code
    ) -> torch.FloatTensor:
        residual = hidden_states

        args = () if USE_PEFT_BACKEND else (scale,)

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, key_tokens, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        attention_mask = attn.prepare_attention_mask(attention_mask, key_tokens, batch_size)
        if attention_mask is not None:
            # expand our mask's singleton query_tokens dimension:
            #   [batch*heads,            1, key_tokens] ->
            #   [batch*heads, query_tokens, key_tokens]
            # so that it can be added as a bias onto the attention scores that xformers computes:
            #   [batch*heads, query_tokens, key_tokens]
            # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
            _, query_tokens, _ = hidden_states.shape
            attention_mask = attention_mask.expand(-1, query_tokens, -1)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states, *args)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states, *args)
        value = attn.to_v(encoder_hidden_states, *args)

        query = attn.head_to_batch_dim(query).contiguous()
        key = attn.head_to_batch_dim(key).contiguous()
        value = attn.head_to_batch_dim(value).contiguous()

        hidden_states = xformers.ops.memory_efficient_attention(
            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
        )
        hidden_states = hidden_states.to(query.dtype)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states, *args)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class PoseAdaptorAttnProcessor(nn.Module):
    def __init__(
            self,
            hidden_size,  # dimension of hidden state
            pose_feature_dim=None,  # dimension of the pose feature
            cross_attention_dim=None,  # dimension of the text embedding
            query_condition=False,
            key_value_condition=False,
            scale=1.0,
        ):
        super().__init__()

        self.hidden_size = hidden_size
        self.pose_feature_dim = pose_feature_dim
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.query_condition = query_condition
        self.key_value_condition = key_value_condition
        assert hidden_size == pose_feature_dim
        if self.query_condition and self.key_value_condition:
            self.qkv_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.qkv_merge.weight)
            init.zeros_(self.qkv_merge.bias)
        elif self.query_condition:
            self.q_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.q_merge.weight)
            init.zeros_(self.q_merge.bias)
        else:
            self.kv_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.kv_merge.weight)
            init.zeros_(self.kv_merge.bias)

    def forward(
            self,
            attn,
            hidden_states,
            pose_feature,
            encoder_hidden_states=None,
            attention_mask=None,
            temb=None,
            scale=None,
        ):
        assert pose_feature is not None
        pose_embedding_scale = (scale or self.scale)

        residual = hidden_states
        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        assert hidden_states.ndim == 3 and pose_feature.ndim == 3

        if self.query_condition and self.key_value_condition:
            assert encoder_hidden_states is None

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states

        assert encoder_hidden_states.ndim == 3

        batch_size, ehs_sequence_length, _ = encoder_hidden_states.shape
        attention_mask = attn.prepare_attention_mask(attention_mask, ehs_sequence_length, batch_size)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        if attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        if self.query_condition and self.key_value_condition:  # only self attention
            query_hidden_state = self.qkv_merge(hidden_states + pose_feature) * pose_embedding_scale + hidden_states
            key_value_hidden_state = query_hidden_state
        elif self.query_condition:
            query_hidden_state = self.q_merge(hidden_states + pose_feature) * pose_embedding_scale + hidden_states
            key_value_hidden_state = encoder_hidden_states
        else:
            key_value_hidden_state = self.kv_merge(encoder_hidden_states + pose_feature) * pose_embedding_scale + encoder_hidden_states
            query_hidden_state = hidden_states

        # original attention
        query = attn.to_q(query_hidden_state)
        key = attn.to_k(key_value_hidden_state)
        value = attn.to_v(key_value_hidden_state)

        query = attn.head_to_batch_dim(query)
        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        attention_probs = attn.get_attention_scores(query, key, attention_mask)
        hidden_states = torch.bmm(attention_probs, value)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class PoseAdaptorAttnProcessor2_0(nn.Module):
    def __init__(
            self,
            hidden_size,  # dimension of hidden state
            pose_feature_dim=None,  # dimension of the pose feature
            cross_attention_dim=None,  # dimension of the text embedding
            query_condition=False,
            key_value_condition=False,
            scale=1.0,
        ):
        super().__init__()
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

        self.hidden_size = hidden_size
        self.pose_feature_dim = pose_feature_dim
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.query_condition = query_condition
        self.key_value_condition = key_value_condition
        assert hidden_size == pose_feature_dim
        if self.query_condition and self.key_value_condition:
            self.qkv_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.qkv_merge.weight)
            init.zeros_(self.qkv_merge.bias)
        elif self.query_condition:
            self.q_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.q_merge.weight)
            init.zeros_(self.q_merge.bias)
        else:
            self.kv_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.kv_merge.weight)
            init.zeros_(self.kv_merge.bias)

    def forward(
            self,
            attn,
            hidden_states,
            pose_feature,
            encoder_hidden_states=None,
            attention_mask=None,
            temb=None,
            scale=None,
        ):
        assert pose_feature is not None
        pose_embedding_scale = (scale or self.scale)

        residual = hidden_states
        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        assert hidden_states.ndim == 3 and pose_feature.ndim == 3

        if self.query_condition and self.key_value_condition:
            assert encoder_hidden_states is None

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states

        assert encoder_hidden_states.ndim == 3

        batch_size, ehs_sequence_length, _ = encoder_hidden_states.shape
        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(attention_mask, ehs_sequence_length, batch_size)
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        if attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        if self.query_condition and self.key_value_condition:  # only self attention
            query_hidden_state = self.qkv_merge(hidden_states + pose_feature) * pose_embedding_scale + hidden_states
            key_value_hidden_state = query_hidden_state
        elif self.query_condition:
            query_hidden_state = self.q_merge(hidden_states + pose_feature) * pose_embedding_scale + hidden_states
            key_value_hidden_state = encoder_hidden_states
        else:
            key_value_hidden_state = self.kv_merge(encoder_hidden_states + pose_feature) * pose_embedding_scale + encoder_hidden_states
            query_hidden_state = hidden_states

        # original attention
        query = attn.to_q(query_hidden_state)
        key = attn.to_k(key_value_hidden_state)
        value = attn.to_v(key_value_hidden_state)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)        # [bs, seq_len, nhead, head_dim] -> [bs, nhead, seq_len, head_dim]
        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        hidden_states = F.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False)  # [bs, nhead, seq_len, head_dim]
        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)    # [bs, seq_len, dim]
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class PoseAdaptorXFormersAttnProcessor(nn.Module):
    def __init__(
            self,
            hidden_size,  # dimension of hidden state
            pose_feature_dim=None,  # dimension of the pose feature
            cross_attention_dim=None,  # dimension of the text embedding
            query_condition=False,
            key_value_condition=False,
            scale=1.0,
            attention_op: Optional[Callable] = None,
        ):
        super().__init__()

        self.hidden_size = hidden_size
        self.pose_feature_dim = pose_feature_dim
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.query_condition = query_condition
        self.key_value_condition = key_value_condition
        self.attention_op = attention_op
        assert hidden_size == pose_feature_dim
        if self.query_condition and self.key_value_condition:
            self.qkv_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.qkv_merge.weight)
            init.zeros_(self.qkv_merge.bias)
        elif self.query_condition:
            self.q_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.q_merge.weight)
            init.zeros_(self.q_merge.bias)
        else:
            self.kv_merge = nn.Linear(hidden_size, hidden_size)
            init.zeros_(self.kv_merge.weight)
            init.zeros_(self.kv_merge.bias)

    def forward(
            self,
            attn,
            hidden_states,
            pose_feature,
            encoder_hidden_states=None,
            attention_mask=None,
            temb=None,
            scale=None,
        ):
        assert pose_feature is not None
        pose_embedding_scale = (scale or self.scale)

        residual = hidden_states
        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        assert hidden_states.ndim == 3 and pose_feature.ndim == 3

        if self.query_condition and self.key_value_condition:
            assert encoder_hidden_states is None

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states

        assert encoder_hidden_states.ndim == 3

        batch_size, ehs_sequence_length, _ = encoder_hidden_states.shape
        attention_mask = attn.prepare_attention_mask(attention_mask, ehs_sequence_length, batch_size)
        if attention_mask is not None:
            # expand our mask's singleton query_tokens dimension:
            #   [batch*heads,            1, key_tokens] ->
            #   [batch*heads, query_tokens, key_tokens]
            # so that it can be added as a bias onto the attention scores that xformers computes:
            #   [batch*heads, query_tokens, key_tokens]
            # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
            _, query_tokens, _ = hidden_states.shape
            attention_mask = attention_mask.expand(-1, query_tokens, -1)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        if attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        if self.query_condition and self.key_value_condition:  # only self attention
            query_hidden_state = self.qkv_merge(hidden_states + pose_feature) * pose_embedding_scale + hidden_states
            key_value_hidden_state = query_hidden_state
        elif self.query_condition:
            query_hidden_state = self.q_merge(hidden_states + pose_feature) * pose_embedding_scale + hidden_states
            key_value_hidden_state = encoder_hidden_states
        else:
            key_value_hidden_state = self.kv_merge(encoder_hidden_states + pose_feature) * pose_embedding_scale + encoder_hidden_states
            query_hidden_state = hidden_states

        # original attention
        query = attn.to_q(query_hidden_state)
        key = attn.to_k(key_value_hidden_state)
        value = attn.to_v(key_value_hidden_state)

        query = attn.head_to_batch_dim(query).contiguous()
        key = attn.head_to_batch_dim(key).contiguous()
        value = attn.head_to_batch_dim(value).contiguous()

        hidden_states = xformers.ops.memory_efficient_attention(
            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
        )
        hidden_states = hidden_states.to(query.dtype)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states