File size: 25,175 Bytes
ce91ea1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Optional

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


class Attention(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.

        upcast_attention (`bool`, *optional*, defaults to False):

            Set to `True` to upcast the attention computation to `float32`.

        upcast_softmax (`bool`, *optional*, defaults to False):

            Set to `True` to upcast the softmax computation to `float32`.

        cross_attention_norm (`str`, *optional*, defaults to `None`):

            The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.

        cross_attention_norm_num_groups (`int`, *optional*, defaults to 32):

            The number of groups to use for the group norm in the cross attention.

        added_kv_proj_dim (`int`, *optional*, defaults to `None`):

            The number of channels to use for the added key and value projections. If `None`, no projection is used.

        norm_num_groups (`int`, *optional*, defaults to `None`):

            The number of groups to use for the group norm in the attention.

        spatial_norm_dim (`int`, *optional*, defaults to `None`):

            The number of channels to use for the spatial normalization.

        out_bias (`bool`, *optional*, defaults to `True`):

            Set to `True` to use a bias in the output linear layer.

        scale_qk (`bool`, *optional*, defaults to `True`):

            Set to `True` to scale the query and key by `1 / sqrt(dim_head)`.

        only_cross_attention (`bool`, *optional*, defaults to `False`):

            Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if

            `added_kv_proj_dim` is not `None`.

        eps (`float`, *optional*, defaults to 1e-5):

            An additional value added to the denominator in group normalization that is used for numerical stability.

        rescale_output_factor (`float`, *optional*, defaults to 1.0):

            A factor to rescale the output by dividing it with this value.

        residual_connection (`bool`, *optional*, defaults to `False`):

            Set to `True` to add the residual connection to the output.

        _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`):

            Set to `True` if the attention block is loaded from a deprecated state dict.

        processor (`AttnProcessor`, *optional*, defaults to `None`):

            The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and

            `AttnProcessor` otherwise.

    """

    def __init__(

        self,

        query_dim: int,

        cross_attention_dim: Optional[int] = None,

        heads: int = 8,

        dim_head: int = 64,

        dropout: float = 0.0,

        bias: bool = False,

        upcast_attention: bool = False,

        upcast_softmax: bool = False,

        cross_attention_norm: Optional[str] = None,

        cross_attention_norm_num_groups: int = 32,

        added_kv_proj_dim: Optional[int] = None,

        norm_num_groups: Optional[int] = None,

        out_bias: bool = True,

        scale_qk: bool = True,

        only_cross_attention: bool = False,

        eps: float = 1e-5,

        rescale_output_factor: float = 1.0,

        residual_connection: bool = False,

        _from_deprecated_attn_block: bool = False,

        processor: Optional["AttnProcessor"] = None,

        out_dim: int = None,

    ):
        super().__init__()
        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
        self.query_dim = query_dim
        self.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.rescale_output_factor = rescale_output_factor
        self.residual_connection = residual_connection
        self.dropout = dropout
        self.fused_projections = False
        self.out_dim = out_dim if out_dim is not None else query_dim

        # we make use of this private variable to know whether this class is loaded
        # with an deprecated state dict so that we can convert it on the fly
        self._from_deprecated_attn_block = _from_deprecated_attn_block

        self.scale_qk = scale_qk
        self.scale = dim_head**-0.5 if self.scale_qk else 1.0

        self.heads = out_dim // dim_head if out_dim is not None else 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.added_kv_proj_dim = added_kv_proj_dim
        self.only_cross_attention = only_cross_attention

        if self.added_kv_proj_dim is None and self.only_cross_attention:
            raise ValueError(
                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
            )

        if norm_num_groups is not None:
            self.group_norm = nn.GroupNorm(
                num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True
            )
        else:
            self.group_norm = None

        self.spatial_norm = None

        if cross_attention_norm is None:
            self.norm_cross = None
        elif cross_attention_norm == "layer_norm":
            self.norm_cross = nn.LayerNorm(self.cross_attention_dim)
        elif cross_attention_norm == "group_norm":
            if self.added_kv_proj_dim is not None:
                # The given `encoder_hidden_states` are initially of shape
                # (batch_size, seq_len, added_kv_proj_dim) before being projected
                # to (batch_size, seq_len, cross_attention_dim). The norm is applied
                # before the projection, so we need to use `added_kv_proj_dim` as
                # the number of channels for the group norm.
                norm_cross_num_channels = added_kv_proj_dim
            else:
                norm_cross_num_channels = self.cross_attention_dim

            self.norm_cross = nn.GroupNorm(
                num_channels=norm_cross_num_channels,
                num_groups=cross_attention_norm_num_groups,
                eps=1e-5,
                affine=True,
            )
        else:
            raise ValueError(
                f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'"
            )

        linear_cls = nn.Linear

        self.linear_cls = linear_cls
        self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias)

        if not self.only_cross_attention:
            # only relevant for the `AddedKVProcessor` classes
            self.to_k = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
            self.to_v = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
        else:
            self.to_k = None
            self.to_v = None

        if self.added_kv_proj_dim is not None:
            self.add_k_proj = linear_cls(added_kv_proj_dim, self.inner_dim)
            self.add_v_proj = linear_cls(added_kv_proj_dim, self.inner_dim)

        self.to_out = nn.ModuleList([])
        self.to_out.append(linear_cls(self.inner_dim, self.out_dim, bias=out_bias))
        self.to_out.append(nn.Dropout(dropout))

        # set attention processor
        # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
        # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
        # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
        if processor is None:
            processor = (
                AttnProcessor2_0()
                if hasattr(F, "scaled_dot_product_attention") and self.scale_qk
                else AttnProcessor()
            )
        self.set_processor(processor)

    def set_processor(self, processor: "AttnProcessor") -> None:
        self.processor = processor

    def forward(

        self,

        hidden_states: torch.FloatTensor,

        encoder_hidden_states: Optional[torch.FloatTensor] = None,

        attention_mask: Optional[torch.FloatTensor] = None,

        **cross_attention_kwargs,

    ) -> torch.Tensor:
        r"""

        The forward method of the `Attention` class.



        Args:

            hidden_states (`torch.Tensor`):

                The hidden states of the query.

            encoder_hidden_states (`torch.Tensor`, *optional*):

                The hidden states of the encoder.

            attention_mask (`torch.Tensor`, *optional*):

                The attention mask to use. If `None`, no mask is applied.

            **cross_attention_kwargs:

                Additional keyword arguments to pass along to the cross attention.



        Returns:

            `torch.Tensor`: The output of the attention layer.

        """
        # The `Attention` class can call different attention processors / attention functions
        # here we simply pass along all tensors to the selected processor class
        # For standard processors that are defined here, `**cross_attention_kwargs` is empty
        return self.processor(
            self,
            hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )

    def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
        r"""

        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads`

        is the number of heads initialized while constructing the `Attention` class.



        Args:

            tensor (`torch.Tensor`): The tensor to reshape.



        Returns:

            `torch.Tensor`: The reshaped tensor.

        """
        head_size = self.heads
        batch_size, seq_len, dim = tensor.shape
        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 head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:
        r"""

        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is

        the number of heads initialized while constructing the `Attention` class.



        Args:

            tensor (`torch.Tensor`): The tensor to reshape.

            out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is

                reshaped to `[batch_size * heads, seq_len, dim // heads]`.



        Returns:

            `torch.Tensor`: The reshaped tensor.

        """
        head_size = self.heads
        batch_size, seq_len, dim = tensor.shape
        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
        tensor = tensor.permute(0, 2, 1, 3)

        if out_dim == 3:
            tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size)

        return tensor

    def get_attention_scores(

        self,

        query: torch.Tensor,

        key: torch.Tensor,

        attention_mask: torch.Tensor = None,

    ) -> torch.Tensor:
        r"""

        Compute the attention scores.



        Args:

            query (`torch.Tensor`): The query tensor.

            key (`torch.Tensor`): The key tensor.

            attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.



        Returns:

            `torch.Tensor`: The attention probabilities/scores.

        """
        dtype = query.dtype
        if self.upcast_attention:
            query = query.float()
            key = key.float()

        if attention_mask is None:
            baddbmm_input = torch.empty(
                query.shape[0],
                query.shape[1],
                key.shape[1],
                dtype=query.dtype,
                device=query.device,
            )
            beta = 0
        else:
            baddbmm_input = attention_mask
            beta = 1

        attention_scores = torch.baddbmm(
            baddbmm_input,
            query,
            key.transpose(-1, -2),
            beta=beta,
            alpha=self.scale,
        )
        del baddbmm_input

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

        attention_probs = attention_scores.softmax(dim=-1)
        del attention_scores

        attention_probs = attention_probs.to(dtype)

        return attention_probs

    def prepare_attention_mask(

        self,

        attention_mask: torch.Tensor,

        target_length: int,

        batch_size: int,

        out_dim: int = 3,

    ) -> torch.Tensor:
        r"""

        Prepare the attention mask for the attention computation.



        Args:

            attention_mask (`torch.Tensor`):

                The attention mask to prepare.

            target_length (`int`):

                The target length of the attention mask. This is the length of the attention mask after padding.

            batch_size (`int`):

                The batch size, which is used to repeat the attention mask.

            out_dim (`int`, *optional*, defaults to `3`):

                The output dimension of the attention mask. Can be either `3` or `4`.



        Returns:

            `torch.Tensor`: The prepared attention mask.

        """
        head_size = self.heads
        if attention_mask is None:
            return attention_mask

        current_length: int = attention_mask.shape[-1]
        if current_length != target_length:
            if attention_mask.device.type == "mps":
                # HACK: MPS: Does not support padding by greater than dimension of input tensor.
                # Instead, we can manually construct the padding tensor.
                padding_shape = (
                    attention_mask.shape[0],
                    attention_mask.shape[1],
                    target_length,
                )
                padding = torch.zeros(
                    padding_shape,
                    dtype=attention_mask.dtype,
                    device=attention_mask.device,
                )
                attention_mask = torch.cat([attention_mask, padding], dim=2)
            else:
                # TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
                #       we want to instead pad by (0, remaining_length), where remaining_length is:
                #       remaining_length: int = target_length - current_length
                # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)

        if out_dim == 3:
            if attention_mask.shape[0] < batch_size * head_size:
                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
        elif out_dim == 4:
            attention_mask = attention_mask.unsqueeze(1)
            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)

        return attention_mask

    def norm_encoder_hidden_states(

        self, encoder_hidden_states: torch.Tensor

    ) -> torch.Tensor:
        r"""

        Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the

        `Attention` class.



        Args:

            encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.



        Returns:

            `torch.Tensor`: The normalized encoder hidden states.

        """
        assert (
            self.norm_cross is not None
        ), "self.norm_cross must be defined to call self.norm_encoder_hidden_states"

        if isinstance(self.norm_cross, nn.LayerNorm):
            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
        elif isinstance(self.norm_cross, nn.GroupNorm):
            # Group norm norms along the channels dimension and expects
            # input to be in the shape of (N, C, *). In this case, we want
            # to norm along the hidden dimension, so we need to move
            # (batch_size, sequence_length, hidden_size) ->
            # (batch_size, hidden_size, sequence_length)
            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
        else:
            assert False

        return encoder_hidden_states

    @torch.no_grad()
    def fuse_projections(self, fuse=True):
        is_cross_attention = self.cross_attention_dim != self.query_dim
        device = self.to_q.weight.data.device
        dtype = self.to_q.weight.data.dtype

        if not is_cross_attention:
            # fetch weight matrices.
            concatenated_weights = torch.cat(
                [self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data]
            )
            in_features = concatenated_weights.shape[1]
            out_features = concatenated_weights.shape[0]

            # create a new single projection layer and copy over the weights.
            self.to_qkv = self.linear_cls(
                in_features, out_features, bias=False, device=device, dtype=dtype
            )
            self.to_qkv.weight.copy_(concatenated_weights)

        else:
            concatenated_weights = torch.cat(
                [self.to_k.weight.data, self.to_v.weight.data]
            )
            in_features = concatenated_weights.shape[1]
            out_features = concatenated_weights.shape[0]

            self.to_kv = self.linear_cls(
                in_features, out_features, bias=False, device=device, dtype=dtype
            )
            self.to_kv.weight.copy_(concatenated_weights)

        self.fused_projections = fuse


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,

    ) -> torch.Tensor:
        residual = hidden_states

        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)

        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)
        value = attn.to_v(encoder_hidden_states)

        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 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,

    ) -> torch.FloatTensor:
        residual = hidden_states

        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
            )

        query = attn.to_q(hidden_states)

        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)
        value = attn.to_v(encoder_hidden_states)

        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)
        # 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