forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathattention_processor.py
More file actions
695 lines (544 loc) 路 28.1 KB
/
Copy pathattention_processor.py
File metadata and controls
695 lines (544 loc) 路 28.1 KB
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
# Copyright 2023 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.
from typing import Callable, Optional, Union
import torch
import torch.nn.functional as F
from torch import nn
from ..utils import deprecate, logging
from ..utils.import_utils import is_xformers_available
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
if is_xformers_available():
import xformers
import xformers.ops
else:
xformers = None
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.
"""
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,
cross_attention_norm: bool = False,
added_kv_proj_dim: Optional[int] = None,
norm_num_groups: Optional[int] = None,
out_bias: bool = True,
scale_qk: bool = True,
processor: Optional["AttnProcessor"] = 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.cross_attention_norm = cross_attention_norm
self.scale = dim_head**-0.5 if scale_qk else 1.0
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.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
if cross_attention_norm:
self.norm_cross = nn.LayerNorm(cross_attention_dim)
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, 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 scale_qk else AttnProcessor()
)
self.set_processor(processor)
def set_use_memory_efficient_attention_xformers(
self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
):
is_lora = hasattr(self, "processor") and isinstance(
self.processor, (LoRAAttnProcessor, LoRAXFormersAttnProcessor)
)
if use_memory_efficient_attention_xformers:
if self.added_kv_proj_dim is not None:
# TODO(Anton, Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP
# which uses this type of cross attention ONLY because the attention mask of format
# [0, ..., -10.000, ..., 0, ...,] is not supported
raise NotImplementedError(
"Memory efficient attention with `xformers` is currently not supported when"
" `self.added_kv_proj_dim` is defined."
)
elif 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
if is_lora:
processor = LoRAXFormersAttnProcessor(
hidden_size=self.processor.hidden_size,
cross_attention_dim=self.processor.cross_attention_dim,
rank=self.processor.rank,
attention_op=attention_op,
)
processor.load_state_dict(self.processor.state_dict())
processor.to(self.processor.to_q_lora.up.weight.device)
else:
processor = XFormersAttnProcessor(attention_op=attention_op)
else:
if is_lora:
processor = LoRAAttnProcessor(
hidden_size=self.processor.hidden_size,
cross_attention_dim=self.processor.cross_attention_dim,
rank=self.processor.rank,
)
processor.load_state_dict(self.processor.state_dict())
processor.to(self.processor.to_q_lora.up.weight.device)
else:
processor = AttnProcessor()
self.set_processor(processor)
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}.")
if slice_size is not None and self.added_kv_proj_dim is not None:
processor = SlicedAttnAddedKVProcessor(slice_size)
elif slice_size is not None:
processor = SlicedAttnProcessor(slice_size)
elif self.added_kv_proj_dim is not None:
processor = AttnAddedKVProcessor()
else:
processor = AttnProcessor()
self.set_processor(processor)
def set_processor(self, processor: "AttnProcessor"):
# if current processor is in `self._modules` and if passed `processor` is not, we need to
# pop `processor` from `self._modules`
if (
hasattr(self, "processor")
and isinstance(self.processor, torch.nn.Module)
and not isinstance(processor, torch.nn.Module)
):
logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")
self._modules.pop("processor")
self.processor = processor
def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, **cross_attention_kwargs):
# 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):
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):
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).reshape(batch_size * head_size, seq_len, dim // head_size)
return tensor
def get_attention_scores(self, query, key, attention_mask=None):
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,
)
if self.upcast_softmax:
attention_scores = attention_scores.float()
attention_probs = attention_scores.softmax(dim=-1)
attention_probs = attention_probs.to(dtype)
return attention_probs
def prepare_attention_mask(self, attention_mask, target_length, batch_size=None):
if batch_size is None:
deprecate(
"batch_size=None",
"0.0.15",
(
"Not passing the `batch_size` parameter to `prepare_attention_mask` can lead to incorrect"
" attention mask preparation and is deprecated behavior. Please make sure to pass `batch_size` to"
" `prepare_attention_mask` when preparing the attention_mask."
),
)
batch_size = 1
head_size = self.heads
if attention_mask is None:
return attention_mask
if attention_mask.shape[-1] != 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:
attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
if attention_mask.shape[0] < batch_size * head_size:
attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
return attention_mask
class AttnProcessor:
def __call__(
self,
attn: Attention,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
):
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)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.cross_attention_norm:
encoder_hidden_states = attn.norm_cross(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)
return hidden_states
class LoRALinearLayer(nn.Module):
def __init__(self, in_features, out_features, rank=4):
super().__init__()
if rank > min(in_features, out_features):
raise ValueError(f"LoRA rank {rank} must be less or equal than {min(in_features, out_features)}")
self.down = nn.Linear(in_features, rank, bias=False)
self.up = nn.Linear(rank, out_features, bias=False)
nn.init.normal_(self.down.weight, std=1 / rank)
nn.init.zeros_(self.up.weight)
def forward(self, hidden_states):
orig_dtype = hidden_states.dtype
dtype = self.down.weight.dtype
down_hidden_states = self.down(hidden_states.to(dtype))
up_hidden_states = self.up(down_hidden_states)
return up_hidden_states.to(orig_dtype)
class LoRAAttnProcessor(nn.Module):
def __init__(self, hidden_size, cross_attention_dim=None, rank=4):
super().__init__()
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.rank = rank
self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank)
self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank)
self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank)
self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank)
def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None, scale=1.0):
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)
query = attn.to_q(hidden_states) + scale * self.to_q_lora(hidden_states)
query = attn.head_to_batch_dim(query)
encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
key = attn.to_k(encoder_hidden_states) + scale * self.to_k_lora(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states) + scale * self.to_v_lora(encoder_hidden_states)
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) + scale * self.to_out_lora(hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
return hidden_states
class AttnAddedKVProcessor:
def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
residual = hidden_states
hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
batch_size, sequence_length, _ = hidden_states.shape
encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
query = attn.head_to_batch_dim(query)
key = attn.to_k(hidden_states)
value = attn.to_v(hidden_states)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
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)
hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
hidden_states = hidden_states + residual
return hidden_states
class XFormersAttnProcessor:
def __init__(self, attention_op: Optional[Callable] = None):
self.attention_op = attention_op
def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
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)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.cross_attention_norm:
encoder_hidden_states = attn.norm_cross(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
query = attn.head_to_batch_dim(query).contiguous()
key = attn.head_to_batch_dim(key).contiguous()
value = attn.head_to_batch_dim(value).contiguous()
hidden_states = xformers.ops.memory_efficient_attention(
query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
)
hidden_states = hidden_states.to(query.dtype)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
return hidden_states
class AttnProcessor2_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, encoder_hidden_states=None, attention_mask=None):
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
inner_dim = hidden_states.shape[-1]
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])
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.cross_attention_norm:
encoder_hidden_states = attn.norm_cross(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
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)
return hidden_states
class LoRAXFormersAttnProcessor(nn.Module):
def __init__(self, hidden_size, cross_attention_dim, rank=4, attention_op: Optional[Callable] = None):
super().__init__()
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.rank = rank
self.attention_op = attention_op
self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank)
self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank)
self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank)
self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank)
def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None, scale=1.0):
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)
query = attn.to_q(hidden_states) + scale * self.to_q_lora(hidden_states)
query = attn.head_to_batch_dim(query).contiguous()
encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
key = attn.to_k(encoder_hidden_states) + scale * self.to_k_lora(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states) + scale * self.to_v_lora(encoder_hidden_states)
key = attn.head_to_batch_dim(key).contiguous()
value = attn.head_to_batch_dim(value).contiguous()
hidden_states = xformers.ops.memory_efficient_attention(
query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states) + scale * self.to_out_lora(hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
return hidden_states
class SlicedAttnProcessor:
def __init__(self, slice_size):
self.slice_size = slice_size
def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
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)
query = attn.to_q(hidden_states)
dim = query.shape[-1]
query = attn.head_to_batch_dim(query)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.cross_attention_norm:
encoder_hidden_states = attn.norm_cross(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
batch_size_attention, query_tokens, _ = query.shape
hidden_states = torch.zeros(
(batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
)
for i in range(batch_size_attention // self.slice_size):
start_idx = i * self.slice_size
end_idx = (i + 1) * self.slice_size
query_slice = query[start_idx:end_idx]
key_slice = key[start_idx:end_idx]
attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
hidden_states[start_idx:end_idx] = attn_slice
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)
return hidden_states
class SlicedAttnAddedKVProcessor:
def __init__(self, slice_size):
self.slice_size = slice_size
def __call__(self, attn: "Attention", hidden_states, encoder_hidden_states=None, attention_mask=None):
residual = hidden_states
hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
batch_size, sequence_length, _ = hidden_states.shape
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
dim = query.shape[-1]
query = attn.head_to_batch_dim(query)
key = attn.to_k(hidden_states)
value = attn.to_v(hidden_states)
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
batch_size_attention, query_tokens, _ = query.shape
hidden_states = torch.zeros(
(batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
)
for i in range(batch_size_attention // self.slice_size):
start_idx = i * self.slice_size
end_idx = (i + 1) * self.slice_size
query_slice = query[start_idx:end_idx]
key_slice = key[start_idx:end_idx]
attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
hidden_states[start_idx:end_idx] = attn_slice
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)
hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
hidden_states = hidden_states + residual
return hidden_states
AttentionProcessor = Union[
AttnProcessor,
XFormersAttnProcessor,
SlicedAttnProcessor,
AttnAddedKVProcessor,
SlicedAttnAddedKVProcessor,
LoRAAttnProcessor,
LoRAXFormersAttnProcessor,
]