File size: 23,614 Bytes
be791d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch import nn
from typing import Optional
from dataclasses import dataclass
from diffusers.utils import BaseOutput
from diffusers.utils.import_utils import is_xformers_available
import torch.nn.functional as F
from einops import rearrange, repeat
import math

@dataclass
class Transformer3DModelOutput(BaseOutput):
    sample: torch.FloatTensor


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

def exists(x):
    return x is not None

class CrossAttention(nn.Module):
    r"""
    copy from diffuser 0.11.1
    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,
        use_relative_position: bool = False,
    ):
        super().__init__()
        # print('num head', heads)
        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
        self.dim_head = dim_head
        # 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 # No use xformers for temporal attention
        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 reshape_for_scores(self, tensor):
        # split heads and dims
        # tensor should be [b (h w)] f (d nd)
        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).contiguous()
        return tensor
    
    def same_batch_dim_to_heads(self, tensor):
        batch_size, head_size, seq_len, dim = tensor.shape # [b (h w)] nd f d
        tensor = tensor.reshape(batch_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, use_image_num=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) # [b (h w)] f (nd * d)

        # print('before reshpape query shape', query.shape)
        dim = query.shape[-1]
        query = self.reshape_heads_to_batch_dim(query) # [b (h w) nd] f d
        # print('after reshape query shape', query.shape)

        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)

        # do not use xformers for temporal attention
        # # 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)
        hidden_states = self._attention(query, key, value, 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,
        )

        # print('query shape', query.shape)
        # print('key shape', key.shape)
        # print('value shape', value.shape)

        if attention_mask is not None:
            # print('attention_mask', attention_mask.shape)
            # print('attention_scores', attention_scores.shape)
            # exit()
            attention_scores = attention_scores + attention_mask

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

        attention_probs = attention_scores.softmax(dim=-1)
        # print(attention_probs.shape)

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

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

        # reshape hidden_states
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        # print(hidden_states.shape)
        # exit()
        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):
        # TODO attention_mask
        query = query.contiguous()
        key = key.contiguous()
        value = value.contiguous()
        # print(query.shape)
        # print(key.shape)
        # print(value.shape)
        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
        # print(hidden_states.shape)
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        # print(hidden_states.shape)
        # exit()
        return hidden_states

class TemporalAttention(CrossAttention):
    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,
                rotary_emb=None):
        super().__init__(query_dim, cross_attention_dim, heads, dim_head, dropout, bias, upcast_attention, upcast_softmax, added_kv_proj_dim, norm_num_groups)
        # relative time positional embeddings
        self.time_rel_pos_bias = RelativePositionBias(heads=heads, max_distance=32) # realistically will not be able to generate that many frames of video... yet
        self.rotary_emb = rotary_emb
        # self.rotary_emb = RotaryEmbedding(32)

    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
        time_rel_pos_bias = self.time_rel_pos_bias(hidden_states.shape[1], device=hidden_states.device)
        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) # [b (h w)] f (nd * d)
        dim = query.shape[-1]
        
        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)
            
        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)

        # Do not use xformers for temporal attention
        # 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, time_rel_pos_bias)
        #     else:
        #         hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask)

        if self._slice_size is None or query.shape[0] // self._slice_size == 1:
            hidden_states = self._attention(query, key, value, attention_mask, time_rel_pos_bias)
        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, time_rel_pos_bias=None):
        if self.upcast_attention:
            query = query.float()
            key = key.float()

        # print('query shape', query.shape)
        # print('key shape', key.shape)
        # print('value shape', value.shape)
        # reshape for adding time positional bais
        query = self.scale * rearrange(query, 'b f (h d) -> b h f d', h=self.heads) # d: dim_head; n: heads
        key = rearrange(key, 'b f (h d) -> b h f d', h=self.heads) # d: dim_head; n: heads
        value = rearrange(value, 'b f (h d) -> b h f d', h=self.heads) # d: dim_head; n: heads
        # print('query shape', query.shape)
        # print('key shape', key.shape)
        # print('value shape', value.shape)

        # torch.baddbmm only accepte 3-D tensor
        # https://runebook.dev/zh/docs/pytorch/generated/torch.baddbmm
        # attention_scores = self.scale * torch.matmul(query, key.transpose(-1, -2))
        if exists(self.rotary_emb):
            query = self.rotary_emb.rotate_queries_or_keys(query)
            key = self.rotary_emb.rotate_queries_or_keys(key)

        attention_scores = torch.einsum('... h i d, ... h j d -> ... h i j', query, key)
        # print('attention_scores shape', attention_scores.shape)
        # print('time_rel_pos_bias shape', time_rel_pos_bias.shape)
        # print('attention_mask shape', attention_mask.shape)

        attention_scores = attention_scores + time_rel_pos_bias
        # print(attention_scores.shape)

        # bert from huggin face
        # attention_scores = attention_scores / math.sqrt(self.dim_head)

        # # Normalize the attention scores to probabilities.
        # attention_probs = nn.functional.softmax(attention_scores, dim=-1)

        if attention_mask is not None:
            # add attention mask
            attention_scores = attention_scores + attention_mask

        # vdm 
        attention_scores = attention_scores - attention_scores.amax(dim = -1, keepdim = True).detach()

        # # Mask out future positions (causal mask)
        # mask = torch.triu(torch.ones(16, 16), diagonal=1).to(device=attention_scores.device, dtype=attention_scores.dtype) # 
        # attention_scores.masked_fill_(mask == 1, float('-inf'))

        # # # disable the fisrt frame
        # mask = torch.zeros(16, 16).to(device=attention_scores.device, dtype=attention_scores.dtype)
        # mask[:, :1] = 1
        # mask[0, 0] = 0
        # attention_scores.masked_fill_(mask == 1, float('-inf'))

        # only enable the first frame to internact with others frames
        # mask = torch.zeros(16, 16).to(device=attention_scores.device, dtype=attention_scores.dtype)
        # mask[:1, 1:] = 1
        # attention_scores.masked_fill_(mask == 1, float('-inf'))

        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
        # print(attention_probs[0][0])

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

        # compute attention output 
        # hidden_states = torch.matmul(attention_probs, value)
        hidden_states = torch.einsum('... h i j, ... h j d -> ... h i d', attention_probs, value)
        # print(hidden_states.shape)
        # hidden_states = self.same_batch_dim_to_heads(hidden_states)
        hidden_states = rearrange(hidden_states, 'b h f d -> b f (h d)')
        # print(hidden_states.shape)
        # exit() 
        return hidden_states
    
class RelativePositionBias(nn.Module):
    def __init__(
        self,
        heads=8,
        num_buckets=32,
        max_distance=128,
    ):
        super().__init__()
        self.num_buckets = num_buckets
        self.max_distance = max_distance
        self.relative_attention_bias = nn.Embedding(num_buckets, heads)

    @staticmethod
    def _relative_position_bucket(relative_position, num_buckets=32, max_distance=128):
        ret = 0
        n = -relative_position

        num_buckets //= 2
        ret += (n < 0).long() * num_buckets
        n = torch.abs(n)

        max_exact = num_buckets // 2
        is_small = n < max_exact

        val_if_large = max_exact + (
            torch.log(n.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact)
        ).long()
        val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1))

        ret += torch.where(is_small, n, val_if_large)
        return ret

    def forward(self, n, device):
        q_pos = torch.arange(n, dtype = torch.long, device = device)
        k_pos = torch.arange(n, dtype = torch.long, device = device)
        rel_pos = rearrange(k_pos, 'j -> 1 j') - rearrange(q_pos, 'i -> i 1')
        rp_bucket = self._relative_position_bucket(rel_pos, num_buckets = self.num_buckets, max_distance = self.max_distance)
        values = self.relative_attention_bias(rp_bucket)
        return rearrange(values, 'i j h -> h i j') # num_heads, num_frames, num_frames
    
class PseudoCrossAttention(CrossAttention):
    def forward(self, hidden_states, encoder_hidden_states=None, base_content=None, attention_mask=None, video_length=None):
        batch_size, sequence_length, _ = hidden_states.shape
        video_length = 17

        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:
            raise NotImplementedError

        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 = rearrange(key, "(b f) d c -> b f d c", f=video_length).contiguous()
        key[:, 1:] = key[:, 1:] + key[:, :1]
        key = rearrange(key, "b f d c -> (b f) d c").contiguous()

        value = rearrange(value, "(b f) d c -> b f d c", f=video_length).contiguous()
        value[:, 1:] = value[:, 1:] + value[:, :1]
        value = rearrange(value, "b f d c -> (b f) d c").contiguous()

        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)

        # hidden_states = rearrange(hidden_states, '(b f) d c -> b f d c', f=video_length).contiguous()
        # hidden_states[:, :1, ...] = base_content
        # hidden_states = rearrange(hidden_states, 'b f d c -> (b f) d c')

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

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