File size: 43,090 Bytes
26555ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from dataclasses import dataclass
from typing import Optional

import torch
import torch.nn.functional as F
from torch import nn

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.embeddings import ImagePositionalEmbeddings
from diffusers.utils import BaseOutput
from diffusers.utils.import_utils import is_xformers_available
from diffusers.models.attention import FeedForward, AdaLayerNorm
from diffusers.models.attention_processor import Attention


@dataclass
class Transformer2DModelOutput(BaseOutput):
    """
    Args:
        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
            Hidden states conditioned on `encoder_hidden_states` input. If discrete, returns probability distributions
            for the unnoised latent pixels.
    """

    sample: torch.FloatTensor


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


class Transformer2DModel(ModelMixin, ConfigMixin):
    """
    Transformer model for image-like data. Takes either discrete (classes of vector embeddings) or continuous (actual
    embeddings) inputs.

    When input is continuous: First, project the input (aka embedding) and reshape to b, t, d. Then apply standard
    transformer action. Finally, reshape to image.

    When input is discrete: First, input (classes of latent pixels) is converted to embeddings and has positional
    embeddings applied, see `ImagePositionalEmbeddings`. Then apply standard transformer action. Finally, predict
    classes of unnoised image.

    Note that it is assumed one of the input classes is the masked latent pixel. The predicted classes of the unnoised
    image do not contain a prediction for the masked pixel as the unnoised image cannot be masked.

    Parameters:
        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
        in_channels (`int`, *optional*):
            Pass if the input is continuous. The number of channels in the input and output.
        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
            `ImagePositionalEmbeddings`.
        num_vector_embeds (`int`, *optional*):
            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
            Includes the class for the masked latent pixel.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
            up to but not more than steps than `num_embeds_ada_norm`.
        attention_bias (`bool`, *optional*):
            Configure if the TransformerBlocks' attention should contain a bias parameter.
    """

    @register_to_config
    def __init__(
        self,
        num_attention_heads: int = 16,
        attention_head_dim: int = 88,
        in_channels: Optional[int] = None,
        num_layers: int = 1,
        dropout: float = 0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        sample_size: Optional[int] = None,
        num_vector_embeds: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
    ):
        super().__init__()
        self.use_linear_projection = use_linear_projection
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        inner_dim = num_attention_heads * attention_head_dim

        # 1. Transformer2DModel can process both standard continous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
        # Define whether input is continuous or discrete depending on configuration
        self.is_input_continuous = in_channels is not None
        self.is_input_vectorized = num_vector_embeds is not None

        if self.is_input_continuous and self.is_input_vectorized:
            raise ValueError(
                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
                " sure that either `in_channels` or `num_vector_embeds` is None."
            )
        elif not self.is_input_continuous and not self.is_input_vectorized:
            raise ValueError(
                f"Has to define either `in_channels`: {in_channels} or `num_vector_embeds`: {num_vector_embeds}. Make"
                " sure that either `in_channels` or `num_vector_embeds` is not None."
            )

        # 2. Define input layers
        if self.is_input_continuous:
            self.in_channels = in_channels

            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
            if use_linear_projection:
                self.proj_in = nn.Linear(in_channels, inner_dim)
            else:
                self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
        elif self.is_input_vectorized:
            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"

            self.height = sample_size
            self.width = sample_size
            self.num_vector_embeds = num_vector_embeds
            self.num_latent_pixels = self.height * self.width

            self.latent_image_embedding = ImagePositionalEmbeddings(
                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
            )

        # 3. Define transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    dropout=dropout,
                    cross_attention_dim=cross_attention_dim,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                    attention_bias=attention_bias,
                    only_cross_attention=only_cross_attention,
                    upcast_attention=upcast_attention,
                )
                for d in range(num_layers)
            ]
        )

        # 4. Define output layers
        if self.is_input_continuous:
            if use_linear_projection:
                self.proj_out = nn.Linear(in_channels, inner_dim)
            else:
                self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
        elif self.is_input_vectorized:
            self.norm_out = nn.LayerNorm(inner_dim)
            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)

    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True):
        """
        Args:
            hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
                When continous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
                hidden_states
            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
                self-attention.
            timestep ( `torch.long`, *optional*):
                Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.

        Returns:
            [`~models.attention.Transformer2DModelOutput`] or `tuple`: [`~models.attention.Transformer2DModelOutput`]
            if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample
            tensor.
        """
        # 1. Input
        if self.is_input_continuous:
            batch, channel, height, weight = hidden_states.shape
            residual = hidden_states

            hidden_states = self.norm(hidden_states)
            if not self.use_linear_projection:
                hidden_states = self.proj_in(hidden_states)
                inner_dim = hidden_states.shape[1]
                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
            else:
                inner_dim = hidden_states.shape[1]
                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
                hidden_states = self.proj_in(hidden_states)
        elif self.is_input_vectorized:
            hidden_states = self.latent_image_embedding(hidden_states)

        # 2. Blocks
        for block in self.transformer_blocks:
            hidden_states = block(hidden_states, encoder_hidden_states=encoder_hidden_states, timestep=timestep)

        # 3. Output
        if self.is_input_continuous:
            if not self.use_linear_projection:
                hidden_states = (
                    hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
                )
                hidden_states = self.proj_out(hidden_states)
            else:
                hidden_states = self.proj_out(hidden_states)
                hidden_states = (
                    hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
                )

            output = hidden_states + residual
        elif self.is_input_vectorized:
            hidden_states = self.norm_out(hidden_states)
            logits = self.out(hidden_states)
            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
            logits = logits.permute(0, 2, 1)

            # log(p(x_0))
            output = F.log_softmax(logits.double(), dim=1).float()

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)


class AttentionBlock(nn.Module):
    """
    An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted
    to the N-d case.
    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
    Uses three q, k, v linear layers to compute attention.

    Parameters:
        channels (`int`): The number of channels in the input and output.
        num_head_channels (`int`, *optional*):
            The number of channels in each head. If None, then `num_heads` = 1.
        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for group norm.
        rescale_output_factor (`float`, *optional*, defaults to 1.0): The factor to rescale the output by.
        eps (`float`, *optional*, defaults to 1e-5): The epsilon value to use for group norm.
    """

    # IMPORTANT;TODO(Patrick, William) - this class will be deprecated soon. Do not use it anymore

    def __init__(
        self,
        channels: int,
        num_head_channels: Optional[int] = None,
        norm_num_groups: int = 32,
        rescale_output_factor: float = 1.0,
        eps: float = 1e-5,
    ):
        super().__init__()
        self.channels = channels

        self.num_heads = channels // num_head_channels if num_head_channels is not None else 1
        self.num_head_size = num_head_channels
        self.group_norm = nn.GroupNorm(num_channels=channels, num_groups=norm_num_groups, eps=eps, affine=True)

        # define q,k,v as linear layers
        self.query = nn.Linear(channels, channels)
        self.key = nn.Linear(channels, channels)
        self.value = nn.Linear(channels, channels)

        self.rescale_output_factor = rescale_output_factor
        self.proj_attn = nn.Linear(channels, channels, 1)

        self._use_memory_efficient_attention_xformers = False

    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool, attention_op: None):
        if not is_xformers_available():
            raise ModuleNotFoundError(
                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
                " xformers",
                name="xformers",
            )
        elif not torch.cuda.is_available():
            raise ValueError(
                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
                " available for GPU "
            )
        else:
            try:
                # Make sure we can run the memory efficient attention
                _ = xformers.ops.memory_efficient_attention(
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                )
            except Exception as e:
                raise e
            self._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers

    def reshape_heads_to_batch_dim(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.num_heads
        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
        return tensor

    def reshape_batch_dim_to_heads(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.num_heads
        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
        return tensor

    def forward(self, hidden_states):
        residual = hidden_states
        batch, channel, height, width = hidden_states.shape

        # norm
        hidden_states = self.group_norm(hidden_states)

        hidden_states = hidden_states.view(batch, channel, height * width).transpose(1, 2)

        # proj to q, k, v
        query_proj = self.query(hidden_states)
        key_proj = self.key(hidden_states)
        value_proj = self.value(hidden_states)

        scale = 1 / math.sqrt(self.channels / self.num_heads)

        query_proj = self.reshape_heads_to_batch_dim(query_proj)
        key_proj = self.reshape_heads_to_batch_dim(key_proj)
        value_proj = self.reshape_heads_to_batch_dim(value_proj)

        if self._use_memory_efficient_attention_xformers:
            # Memory efficient attention
            hidden_states = xformers.ops.memory_efficient_attention(query_proj, key_proj, value_proj, attn_bias=None)
            hidden_states = hidden_states.to(query_proj.dtype)
        else:
            attention_scores = torch.baddbmm(
                torch.empty(
                    query_proj.shape[0],
                    query_proj.shape[1],
                    key_proj.shape[1],
                    dtype=query_proj.dtype,
                    device=query_proj.device,
                ),
                query_proj,
                key_proj.transpose(-1, -2),
                beta=0,
                alpha=scale,
            )
            attention_probs = torch.softmax(attention_scores.float(), dim=-1).type(attention_scores.dtype)
            hidden_states = torch.bmm(attention_probs, value_proj)

        # reshape hidden_states
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)

        # compute next hidden_states
        hidden_states = self.proj_attn(hidden_states)

        hidden_states = hidden_states.transpose(-1, -2).reshape(batch, channel, height, width)

        # res connect and rescale
        hidden_states = (hidden_states + residual) / self.rescale_output_factor
        return hidden_states


class BasicTransformerBlock(nn.Module):
    r"""
    A basic Transformer block.

    Parameters:
        dim (`int`): The number of channels in the input and output.
        num_attention_heads (`int`): The number of heads to use for multi-head attention.
        attention_head_dim (`int`): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm (:
            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
        attention_bias (:
            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
    ):
        super().__init__()
        self.only_cross_attention = only_cross_attention
        self.use_ada_layer_norm = num_embeds_ada_norm is not None

        # 1. Self-Attn
        self.attn1 = CrossAttention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
            upcast_attention=upcast_attention,
        )  # is a self-attention
        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)

        # 2. Cross-Attn
        if cross_attention_dim is not None:
            self.attn2 = CrossAttention(
                query_dim=dim,
                cross_attention_dim=cross_attention_dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
            )  # is self-attn if encoder_hidden_states is none
        else:
            self.attn2 = None

        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)

        if cross_attention_dim is not None:
            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
        else:
            self.norm2 = None

        # 3. Feed-forward
        self.norm3 = nn.LayerNorm(dim)

    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool, attention_op: None):
        if not is_xformers_available():
            print("Here is how to install it")
            raise ModuleNotFoundError(
                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
                " xformers",
                name="xformers",
            )
        elif not torch.cuda.is_available():
            raise ValueError(
                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
                " available for GPU "
            )
        else:
            try:
                # Make sure we can run the memory efficient attention
                _ = xformers.ops.memory_efficient_attention(
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                )
            except Exception as e:
                raise e
            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
            if self.attn2 is not None:
                self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers

    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None):
        # 1. Self-Attention
        norm_hidden_states = (
            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
        )

        if self.only_cross_attention:
            hidden_states = (
                self.attn1(norm_hidden_states, encoder_hidden_states, attention_mask=attention_mask) + hidden_states
            )
        else:
            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask) + hidden_states

        if self.attn2 is not None:
            # 2. Cross-Attention
            norm_hidden_states = (
                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
            )
            hidden_states = (
                self.attn2(
                    norm_hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
                )
                + hidden_states
            )

        # 3. Feed-forward
        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states

        return hidden_states


class CrossAttention(nn.Module):
    r"""
    A cross attention layer.

    Parameters:
        query_dim (`int`): The number of channels in the query.
        cross_attention_dim (`int`, *optional*):
            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
        heads (`int`,  *optional*, defaults to 8): The number of heads to use for multi-head attention.
        dim_head (`int`,  *optional*, defaults to 64): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        bias (`bool`, *optional*, defaults to False):
            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
    """

    def __init__(
        self,
        query_dim: int,
        cross_attention_dim: Optional[int] = None,
        heads: int = 8,
        dim_head: int = 64,
        dropout: float = 0.0,
        bias=False,
        upcast_attention: bool = False,
        upcast_softmax: bool = False,
        added_kv_proj_dim: Optional[int] = None,
        norm_num_groups: Optional[int] = None,
    ):
        super().__init__()
        inner_dim = dim_head * heads
        cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
        self.upcast_attention = upcast_attention
        self.upcast_softmax = upcast_softmax

        self.scale = dim_head**-0.5

        self.heads = heads
        # for slice_size > 0 the attention score computation
        # is split across the batch axis to save memory
        # You can set slice_size with `set_attention_slice`
        self.sliceable_head_dim = heads
        self._slice_size = None
        self._use_memory_efficient_attention_xformers = False
        self.added_kv_proj_dim = added_kv_proj_dim

        if norm_num_groups is not None:
            self.group_norm = nn.GroupNorm(num_channels=inner_dim, num_groups=norm_num_groups, eps=1e-5, affine=True)
        else:
            self.group_norm = None

        self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
        self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
        self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)

        if self.added_kv_proj_dim is not None:
            self.add_k_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
            self.add_v_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)

        self.to_out = nn.ModuleList([])
        self.to_out.append(nn.Linear(inner_dim, query_dim))
        self.to_out.append(nn.Dropout(dropout))

    def reshape_heads_to_batch_dim(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
        return tensor

    def reshape_batch_dim_to_heads(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
        return tensor

    def set_attention_slice(self, slice_size):
        if slice_size is not None and slice_size > self.sliceable_head_dim:
            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")

        self._slice_size = slice_size

    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
        batch_size, sequence_length, _ = hidden_states.shape

        encoder_hidden_states = encoder_hidden_states

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

        query = self.to_q(hidden_states)
        dim = query.shape[-1]
        query = self.reshape_heads_to_batch_dim(query)

        if self.added_kv_proj_dim is not None:
            key = self.to_k(hidden_states)
            value = self.to_v(hidden_states)
            encoder_hidden_states_key_proj = self.add_k_proj(encoder_hidden_states)
            encoder_hidden_states_value_proj = self.add_v_proj(encoder_hidden_states)

            key = self.reshape_heads_to_batch_dim(key)
            value = self.reshape_heads_to_batch_dim(value)
            encoder_hidden_states_key_proj = self.reshape_heads_to_batch_dim(encoder_hidden_states_key_proj)
            encoder_hidden_states_value_proj = self.reshape_heads_to_batch_dim(encoder_hidden_states_value_proj)

            key = torch.concat([encoder_hidden_states_key_proj, key], dim=1)
            value = torch.concat([encoder_hidden_states_value_proj, value], dim=1)
        else:
            encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
            key = self.to_k(encoder_hidden_states)
            value = self.to_v(encoder_hidden_states)

            key = self.reshape_heads_to_batch_dim(key)
            value = self.reshape_heads_to_batch_dim(value)

        if attention_mask is not None:
            if attention_mask.shape[-1] != query.shape[1]:
                target_length = query.shape[1]
                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
                attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)

        # attention, what we cannot get enough of
        if self._use_memory_efficient_attention_xformers:
            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
            # Some versions of xformers return output in fp32, cast it back to the dtype of the input
            hidden_states = hidden_states.to(query.dtype)
        else:
            if self._slice_size is None or query.shape[0] // self._slice_size == 1:
                hidden_states = self._attention(query, key, value, attention_mask)
            else:
                hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask)

        # linear proj
        hidden_states = self.to_out[0](hidden_states)

        # dropout
        hidden_states = self.to_out[1](hidden_states)
        return hidden_states

    def _attention(self, query, key, value, attention_mask=None):
        if self.upcast_attention:
            query = query.float()
            key = key.float()

        attention_scores = torch.baddbmm(
            torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device),
            query,
            key.transpose(-1, -2),
            beta=0,
            alpha=self.scale,
        )

        if attention_mask is not None:
            attention_scores = attention_scores + attention_mask

        if self.upcast_softmax:
            attention_scores = attention_scores.float()

        attention_probs = attention_scores.softmax(dim=-1)

        # cast back to the original dtype
        attention_probs = attention_probs.to(value.dtype)

        # compute attention output
        hidden_states = torch.bmm(attention_probs, value)

        # reshape hidden_states
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        return hidden_states

    def _sliced_attention(self, query, key, value, sequence_length, dim, attention_mask):
        batch_size_attention = query.shape[0]
        hidden_states = torch.zeros(
            (batch_size_attention, sequence_length, dim // self.heads), device=query.device, dtype=query.dtype
        )
        slice_size = self._slice_size if self._slice_size is not None else hidden_states.shape[0]
        for i in range(hidden_states.shape[0] // slice_size):
            start_idx = i * slice_size
            end_idx = (i + 1) * slice_size

            query_slice = query[start_idx:end_idx]
            key_slice = key[start_idx:end_idx]

            if self.upcast_attention:
                query_slice = query_slice.float()
                key_slice = key_slice.float()

            attn_slice = torch.baddbmm(
                torch.empty(slice_size, query.shape[1], key.shape[1], dtype=query_slice.dtype, device=query.device),
                query_slice,
                key_slice.transpose(-1, -2),
                beta=0,
                alpha=self.scale,
            )

            if attention_mask is not None:
                attn_slice = attn_slice + attention_mask[start_idx:end_idx]

            if self.upcast_softmax:
                attn_slice = attn_slice.float()

            attn_slice = attn_slice.softmax(dim=-1)

            # cast back to the original dtype
            attn_slice = attn_slice.to(value.dtype)
            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])

            hidden_states[start_idx:end_idx] = attn_slice

        # reshape hidden_states
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        return hidden_states

    def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
        query = query.contiguous()
        key = key.contiguous()
        value = value.contiguous()
        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        return hidden_states


class FeedForward(nn.Module):
    r"""
    A feed-forward layer.

    Parameters:
        dim (`int`): The number of channels in the input.
        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
    """

    def __init__(
        self,
        dim: int,
        dim_out: Optional[int] = None,
        mult: int = 4,
        dropout: float = 0.0,
        activation_fn: str = "geglu",
    ):
        super().__init__()
        inner_dim = int(dim * mult)
        dim_out = dim_out if dim_out is not None else dim

        if activation_fn == "gelu":
            act_fn = GELU(dim, inner_dim)
        elif activation_fn == "geglu":
            act_fn = GEGLU(dim, inner_dim)
        elif activation_fn == "geglu-approximate":
            act_fn = ApproximateGELU(dim, inner_dim)

        self.net = nn.ModuleList([])
        # project in
        self.net.append(act_fn)
        # project dropout
        self.net.append(nn.Dropout(dropout))
        # project out
        self.net.append(nn.Linear(inner_dim, dim_out))

    def forward(self, hidden_states):
        for module in self.net:
            hidden_states = module(hidden_states)
        return hidden_states


class GELU(nn.Module):
    r"""
    GELU activation function
    """

    def __init__(self, dim_in: int, dim_out: int):
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out)

    def gelu(self, gate):
        if gate.device.type != "mps":
            return F.gelu(gate)
        # mps: gelu is not implemented for float16
        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)

    def forward(self, hidden_states):
        hidden_states = self.proj(hidden_states)
        hidden_states = self.gelu(hidden_states)
        return hidden_states


# feedforward
class GEGLU(nn.Module):
    r"""
    A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.

    Parameters:
        dim_in (`int`): The number of channels in the input.
        dim_out (`int`): The number of channels in the output.
    """

    def __init__(self, dim_in: int, dim_out: int):
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out * 2)

    def gelu(self, gate):
        if gate.device.type != "mps":
            return F.gelu(gate)
        # mps: gelu is not implemented for float16
        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)

    def forward(self, hidden_states):
        hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
        return hidden_states * self.gelu(gate)


class ApproximateGELU(nn.Module):
    """
    The approximate form of Gaussian Error Linear Unit (GELU)

    For more details, see section 2: https://arxiv.org/abs/1606.08415
    """

    def __init__(self, dim_in: int, dim_out: int):
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out)

    def forward(self, x):
        x = self.proj(x)
        return x * torch.sigmoid(1.702 * x)


class AdaLayerNorm(nn.Module):
    """
    Norm layer modified to incorporate timestep embeddings.
    """

    def __init__(self, embedding_dim, num_embeddings):
        super().__init__()
        self.emb = nn.Embedding(num_embeddings, embedding_dim)
        self.silu = nn.SiLU()
        self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)

    def forward(self, x, timestep):
        emb = self.linear(self.silu(self.emb(timestep)))
        scale, shift = torch.chunk(emb, 2)
        x = self.norm(x) * (1 + scale) + shift
        return x


class DualTransformer2DModel(nn.Module):
    """
    Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference.

    Parameters:
        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
        in_channels (`int`, *optional*):
            Pass if the input is continuous. The number of channels in the input and output.
        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
        dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
            `ImagePositionalEmbeddings`.
        num_vector_embeds (`int`, *optional*):
            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
            Includes the class for the masked latent pixel.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
            up to but not more than steps than `num_embeds_ada_norm`.
        attention_bias (`bool`, *optional*):
            Configure if the TransformerBlocks' attention should contain a bias parameter.
    """

    def __init__(
        self,
        num_attention_heads: int = 16,
        attention_head_dim: int = 88,
        in_channels: Optional[int] = None,
        num_layers: int = 1,
        dropout: float = 0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        sample_size: Optional[int] = None,
        num_vector_embeds: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
    ):
        super().__init__()
        self.transformers = nn.ModuleList(
            [
                Transformer2DModel(
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                    in_channels=in_channels,
                    num_layers=num_layers,
                    dropout=dropout,
                    norm_num_groups=norm_num_groups,
                    cross_attention_dim=cross_attention_dim,
                    attention_bias=attention_bias,
                    sample_size=sample_size,
                    num_vector_embeds=num_vector_embeds,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                )
                for _ in range(2)
            ]
        )

        # Variables that can be set by a pipeline:

        # The ratio of transformer1 to transformer2's output states to be combined during inference
        self.mix_ratio = 0.5

        # The shape of `encoder_hidden_states` is expected to be
        # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)`
        self.condition_lengths = [77, 257]

        # Which transformer to use to encode which condition.
        # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])`
        self.transformer_index_for_condition = [1, 0]

    def forward(
        self, hidden_states, encoder_hidden_states, timestep=None, attention_mask=None, return_dict: bool = True
    ):
        """
        Args:
            hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
                When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
                hidden_states
            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
                self-attention.
            timestep ( `torch.long`, *optional*):
                Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
            attention_mask (`torch.FloatTensor`, *optional*):
                Optional attention mask to be applied in CrossAttention
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.

        Returns:
            [`~models.attention.Transformer2DModelOutput`] or `tuple`: [`~models.attention.Transformer2DModelOutput`]
            if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample
            tensor.
        """
        input_states = hidden_states

        encoded_states = []
        tokens_start = 0
        # attention_mask is not used yet
        for i in range(2):
            # for each of the two transformers, pass the corresponding condition tokens
            condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]]
            transformer_index = self.transformer_index_for_condition[i]
            encoded_state = self.transformers[transformer_index](
                input_states,
                encoder_hidden_states=condition_state,
                timestep=timestep,
                return_dict=False,
            )[0]
            encoded_states.append(encoded_state - input_states)
            tokens_start += self.condition_lengths[i]

        output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio)
        output_states = output_states + input_states

        if not return_dict:
            return (output_states,)

        return Transformer2DModelOutput(sample=output_states)