Repository navigation
Expand file tree
/
Copy pathattention_processor.py
More file actions
executable file
路5679 lines (4556 loc) 路 234 KB
/
Copy pathattention_processor.py
File metadata and controls
executable file
路5679 lines (4556 loc) 路 234 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
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
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright 2026 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 __future__ import annotations
import inspect
import math
from typing import Callable
import torch
import torch.nn.functional as F
from torch import nn
from ..image_processor import IPAdapterMaskProcessor
from ..utils import deprecate, is_torch_xla_available, logging
from ..utils.import_utils import is_torch_npu_available, is_torch_xla_version, is_xformers_available
from ..utils.torch_utils import is_torch_version, maybe_allow_in_graph
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
if is_torch_npu_available():
import torch_npu
if is_xformers_available():
import xformers
import xformers.ops
else:
xformers = None
if is_torch_xla_available():
# flash attention pallas kernel is introduced in the torch_xla 2.3 release.
if is_torch_xla_version(">", "2.2"):
from torch_xla.experimental.custom_kernel import flash_attention
from torch_xla.runtime import is_spmd
XLA_AVAILABLE = True
else:
XLA_AVAILABLE = False
@maybe_allow_in_graph
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.
kv_heads (`int`, *optional*, defaults to `None`):
The number of key and value heads to use for multi-head attention. Defaults to `heads`. If
`kv_heads=heads`, the model will use Multi Head Attention (MHA), if `kv_heads=1` the model will use Multi
Query Attention (MQA) otherwise GQA is used.
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: int | None = None,
heads: int = 8,
kv_heads: int | None = None,
dim_head: int = 64,
dropout: float = 0.0,
bias: bool = False,
upcast_attention: bool = False,
upcast_softmax: bool = False,
cross_attention_norm: str | None = None,
cross_attention_norm_num_groups: int = 32,
qk_norm: str | None = None,
added_kv_proj_dim: int | None = None,
added_proj_bias: bool | None = True,
norm_num_groups: int | None = None,
spatial_norm_dim: int | None = 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: "AttnProcessor" | None = None,
out_dim: int = None,
out_context_dim: int = None,
context_pre_only=None,
pre_only=False,
elementwise_affine: bool = True,
is_causal: bool = False,
):
super().__init__()
# To prevent circular import.
from .normalization import FP32LayerNorm, LpNorm, RMSNorm
self.inner_dim = out_dim if out_dim is not None else dim_head * heads
self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads
self.query_dim = query_dim
self.use_bias = bias
self.is_cross_attention = cross_attention_dim is not None
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
self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
self.context_pre_only = context_pre_only
self.pre_only = pre_only
self.is_causal = is_causal
# 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
if spatial_norm_dim is not None:
self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim)
else:
self.spatial_norm = None
if qk_norm is None:
self.norm_q = None
self.norm_k = None
elif qk_norm == "layer_norm":
self.norm_q = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
self.norm_k = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
elif qk_norm == "fp32_layer_norm":
self.norm_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
self.norm_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
elif qk_norm == "layer_norm_across_heads":
# Lumina applies qk norm across all heads
self.norm_q = nn.LayerNorm(dim_head * heads, eps=eps)
self.norm_k = nn.LayerNorm(dim_head * kv_heads, eps=eps)
elif qk_norm == "rms_norm":
self.norm_q = RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
self.norm_k = RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
elif qk_norm == "rms_norm_across_heads":
# LTX applies qk norm across all heads
self.norm_q = RMSNorm(dim_head * heads, eps=eps)
self.norm_k = RMSNorm(dim_head * kv_heads, eps=eps)
elif qk_norm == "l2":
self.norm_q = LpNorm(p=2, dim=-1, eps=eps)
self.norm_k = LpNorm(p=2, dim=-1, eps=eps)
else:
raise ValueError(
f"unknown qk_norm: {qk_norm}. Should be one of None, 'layer_norm', 'fp32_layer_norm', 'layer_norm_across_heads', 'rms_norm', 'rms_norm_across_heads', 'l2'."
)
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'"
)
self.to_q = nn.Linear(query_dim, self.inner_dim, bias=bias)
if not self.only_cross_attention:
# only relevant for the `AddedKVProcessor` classes
self.to_k = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
self.to_v = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
else:
self.to_k = None
self.to_v = None
self.added_proj_bias = added_proj_bias
if self.added_kv_proj_dim is not None:
self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
if self.context_pre_only is not None:
self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
else:
self.add_q_proj = None
self.add_k_proj = None
self.add_v_proj = None
if not self.pre_only:
self.to_out = nn.ModuleList([])
self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
self.to_out.append(nn.Dropout(dropout))
else:
self.to_out = None
if self.context_pre_only is not None and not self.context_pre_only:
self.to_add_out = nn.Linear(self.inner_dim, self.out_context_dim, bias=out_bias)
else:
self.to_add_out = None
if qk_norm is not None and added_kv_proj_dim is not None:
if qk_norm == "layer_norm":
self.norm_added_q = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
self.norm_added_k = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
elif qk_norm == "fp32_layer_norm":
self.norm_added_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
self.norm_added_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
elif qk_norm == "rms_norm":
self.norm_added_q = RMSNorm(dim_head, eps=eps)
self.norm_added_k = RMSNorm(dim_head, eps=eps)
elif qk_norm == "rms_norm_across_heads":
# Wan applies qk norm across all heads
# Wan also doesn't apply a q norm
self.norm_added_q = None
self.norm_added_k = RMSNorm(dim_head * kv_heads, eps=eps)
else:
raise ValueError(
f"unknown qk_norm: {qk_norm}. Should be one of `None,'layer_norm','fp32_layer_norm','rms_norm'`"
)
else:
self.norm_added_q = None
self.norm_added_k = None
# 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_use_xla_flash_attention(
self,
use_xla_flash_attention: bool,
partition_spec: tuple[str | None, ...] | None = None,
is_flux=False,
) -> None:
r"""
Set whether to use xla flash attention from `torch_xla` or not.
Args:
use_xla_flash_attention (`bool`):
Whether to use pallas flash attention kernel from `torch_xla` or not.
partition_spec (`tuple[]`, *optional*):
Specify the partition specification if using SPMD. Otherwise None.
"""
if use_xla_flash_attention:
if not is_torch_xla_available:
raise "torch_xla is not available"
elif is_torch_xla_version("<", "2.3"):
raise "flash attention pallas kernel is supported from torch_xla version 2.3"
elif is_spmd() and is_torch_xla_version("<", "2.4"):
raise "flash attention pallas kernel using SPMD is supported from torch_xla version 2.4"
else:
if is_flux:
processor = XLAFluxFlashAttnProcessor2_0(partition_spec)
else:
processor = XLAFlashAttnProcessor2_0(partition_spec)
else:
processor = (
AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
)
self.set_processor(processor)
def set_use_npu_flash_attention(self, use_npu_flash_attention: bool) -> None:
r"""
Set whether to use npu flash attention from `torch_npu` or not.
"""
if use_npu_flash_attention:
processor = AttnProcessorNPU()
else:
# 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
processor = (
AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
)
self.set_processor(processor)
def set_use_memory_efficient_attention_xformers(
self, use_memory_efficient_attention_xformers: bool, attention_op: Callable | None = None
) -> None:
r"""
Set whether to use memory efficient attention from `xformers` or not.
Args:
use_memory_efficient_attention_xformers (`bool`):
Whether to use memory efficient attention from `xformers` or not.
attention_op (`Callable`, *optional*):
The attention operation to use. Defaults to `None` which uses the default attention operation from
`xformers`.
"""
is_custom_diffusion = hasattr(self, "processor") and isinstance(
self.processor,
(CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0),
)
is_added_kv_processor = hasattr(self, "processor") and isinstance(
self.processor,
(
AttnAddedKVProcessor,
AttnAddedKVProcessor2_0,
SlicedAttnAddedKVProcessor,
XFormersAttnAddedKVProcessor,
),
)
is_ip_adapter = hasattr(self, "processor") and isinstance(
self.processor,
(IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, IPAdapterXFormersAttnProcessor),
)
is_joint_processor = hasattr(self, "processor") and isinstance(
self.processor,
(
JointAttnProcessor2_0,
XFormersJointAttnProcessor,
),
)
if use_memory_efficient_attention_xformers:
if is_added_kv_processor and is_custom_diffusion:
raise NotImplementedError(
f"Memory efficient attention is currently not supported for custom diffusion for attention processor type {self.processor}"
)
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
dtype = None
if attention_op is not None:
op_fw, op_bw = attention_op
dtype, *_ = op_fw.SUPPORTED_DTYPES
q = torch.randn((1, 2, 40), device="cuda", dtype=dtype)
_ = xformers.ops.memory_efficient_attention(q, q, q)
except Exception as e:
raise e
if is_custom_diffusion:
processor = CustomDiffusionXFormersAttnProcessor(
train_kv=self.processor.train_kv,
train_q_out=self.processor.train_q_out,
hidden_size=self.processor.hidden_size,
cross_attention_dim=self.processor.cross_attention_dim,
attention_op=attention_op,
)
processor.load_state_dict(self.processor.state_dict())
if hasattr(self.processor, "to_k_custom_diffusion"):
processor.to(self.processor.to_k_custom_diffusion.weight.device)
elif is_added_kv_processor:
# TODO(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
# throw warning
logger.info(
"Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation."
)
processor = XFormersAttnAddedKVProcessor(attention_op=attention_op)
elif is_ip_adapter:
processor = IPAdapterXFormersAttnProcessor(
hidden_size=self.processor.hidden_size,
cross_attention_dim=self.processor.cross_attention_dim,
num_tokens=self.processor.num_tokens,
scale=self.processor.scale,
attention_op=attention_op,
)
processor.load_state_dict(self.processor.state_dict())
if hasattr(self.processor, "to_k_ip"):
processor.to(
device=self.processor.to_k_ip[0].weight.device, dtype=self.processor.to_k_ip[0].weight.dtype
)
elif is_joint_processor:
processor = XFormersJointAttnProcessor(attention_op=attention_op)
else:
processor = XFormersAttnProcessor(attention_op=attention_op)
else:
if is_custom_diffusion:
attn_processor_class = (
CustomDiffusionAttnProcessor2_0
if hasattr(F, "scaled_dot_product_attention")
else CustomDiffusionAttnProcessor
)
processor = attn_processor_class(
train_kv=self.processor.train_kv,
train_q_out=self.processor.train_q_out,
hidden_size=self.processor.hidden_size,
cross_attention_dim=self.processor.cross_attention_dim,
)
processor.load_state_dict(self.processor.state_dict())
if hasattr(self.processor, "to_k_custom_diffusion"):
processor.to(self.processor.to_k_custom_diffusion.weight.device)
elif is_ip_adapter:
processor = IPAdapterAttnProcessor2_0(
hidden_size=self.processor.hidden_size,
cross_attention_dim=self.processor.cross_attention_dim,
num_tokens=self.processor.num_tokens,
scale=self.processor.scale,
)
processor.load_state_dict(self.processor.state_dict())
if hasattr(self.processor, "to_k_ip"):
processor.to(
device=self.processor.to_k_ip[0].weight.device, dtype=self.processor.to_k_ip[0].weight.dtype
)
else:
# 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
processor = (
AttnProcessor2_0()
if hasattr(F, "scaled_dot_product_attention") and self.scale_qk
else AttnProcessor()
)
self.set_processor(processor)
def set_attention_slice(self, slice_size: int) -> None:
r"""
Set the slice size for attention computation.
Args:
slice_size (`int`):
The slice size for attention computation.
"""
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:
# 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
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:
r"""
Set the attention processor to use.
Args:
processor (`AttnProcessor`):
The attention processor to use.
"""
# 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 get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor":
r"""
Get the attention processor in use.
Args:
return_deprecated_lora (`bool`, *optional*, defaults to `False`):
Set to `True` to return the deprecated LoRA attention processor.
Returns:
"AttentionProcessor": The attention processor in use.
"""
if not return_deprecated_lora:
return self.processor
def forward(
self,
hidden_states: torch.Tensor,
encoder_hidden_states: torch.Tensor | None = None,
attention_mask: torch.Tensor | None = 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
attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
quiet_attn_parameters = {"ip_adapter_masks", "ip_hidden_states"}
unused_kwargs = [
k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters
]
if len(unused_kwargs) > 0:
logger.warning(
f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
)
cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters}
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
if tensor.ndim == 3:
batch_size, seq_len, dim = tensor.shape
extra_dim = 1
else:
batch_size, extra_dim, seq_len, dim = tensor.shape
tensor = tensor.reshape(batch_size, seq_len * extra_dim, 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 * extra_dim, dim // head_size)
return tensor
def get_attention_scores(
self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor | None = 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, output_size=attention_mask.shape[0] * head_size
)
elif out_dim == 4:
attention_mask = attention_mask.unsqueeze(1)
attention_mask = attention_mask.repeat_interleave(
head_size, dim=1, output_size=attention_mask.shape[1] * head_size
)
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):
device = self.to_q.weight.data.device
dtype = self.to_q.weight.data.dtype
if not self.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 = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
self.to_qkv.weight.copy_(concatenated_weights)
if self.use_bias:
concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
self.to_qkv.bias.copy_(concatenated_bias)
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 = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
self.to_kv.weight.copy_(concatenated_weights)
if self.use_bias:
concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
self.to_kv.bias.copy_(concatenated_bias)
# handle added projections for SD3 and others.
if (
getattr(self, "add_q_proj", None) is not None
and getattr(self, "add_k_proj", None) is not None
and getattr(self, "add_v_proj", None) is not None
):
concatenated_weights = torch.cat(
[self.add_q_proj.weight.data, self.add_k_proj.weight.data, self.add_v_proj.weight.data]
)
in_features = concatenated_weights.shape[1]
out_features = concatenated_weights.shape[0]
self.to_added_qkv = nn.Linear(
in_features, out_features, bias=self.added_proj_bias, device=device, dtype=dtype
)
self.to_added_qkv.weight.copy_(concatenated_weights)
if self.added_proj_bias:
concatenated_bias = torch.cat(
[self.add_q_proj.bias.data, self.add_k_proj.bias.data, self.add_v_proj.bias.data]
)
self.to_added_qkv.bias.copy_(concatenated_bias)
self.fused_projections = fuse
class SanaMultiscaleAttentionProjection(nn.Module):
def __init__(
self,
in_channels: int,
num_attention_heads: int,
kernel_size: int,
) -> None:
super().__init__()
channels = 3 * in_channels
self.proj_in = nn.Conv2d(
channels,
channels,
kernel_size,
padding=kernel_size // 2,
groups=channels,
bias=False,
)
self.proj_out = nn.Conv2d(channels, channels, 1, 1, 0, groups=3 * num_attention_heads, bias=False)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
hidden_states = self.proj_in(hidden_states)
hidden_states = self.proj_out(hidden_states)
return hidden_states
class SanaMultiscaleLinearAttention(nn.Module):
r"""Lightweight multi-scale linear attention"""
def __init__(
self,
in_channels: int,
out_channels: int,
num_attention_heads: int | None = None,
attention_head_dim: int = 8,
mult: float = 1.0,
norm_type: str = "batch_norm",
kernel_sizes: tuple[int, ...] = (5,),
eps: float = 1e-15,
residual_connection: bool = False,
):
super().__init__()
# To prevent circular import
from .normalization import get_normalization
self.eps = eps
self.attention_head_dim = attention_head_dim
self.norm_type = norm_type
self.residual_connection = residual_connection
num_attention_heads = (
int(in_channels // attention_head_dim * mult) if num_attention_heads is None else num_attention_heads
)
inner_dim = num_attention_heads * attention_head_dim
self.to_q = nn.Linear(in_channels, inner_dim, bias=False)
self.to_k = nn.Linear(in_channels, inner_dim, bias=False)
self.to_v = nn.Linear(in_channels, inner_dim, bias=False)
self.to_qkv_multiscale = nn.ModuleList()
for kernel_size in kernel_sizes:
self.to_qkv_multiscale.append(
SanaMultiscaleAttentionProjection(inner_dim, num_attention_heads, kernel_size)
)
self.nonlinearity = nn.ReLU()
self.to_out = nn.Linear(inner_dim * (1 + len(kernel_sizes)), out_channels, bias=False)
self.norm_out = get_normalization(norm_type, num_features=out_channels)
self.processor = SanaMultiscaleAttnProcessor2_0()
def apply_linear_attention(self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor) -> torch.Tensor:
value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1) # Adds padding
scores = torch.matmul(value, key.transpose(-1, -2))
hidden_states = torch.matmul(scores, query)
hidden_states = hidden_states.to(dtype=torch.float32)
hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps)
return hidden_states
def apply_quadratic_attention(self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor) -> torch.Tensor:
scores = torch.matmul(key.transpose(-1, -2), query)
scores = scores.to(dtype=torch.float32)
scores = scores / (torch.sum(scores, dim=2, keepdim=True) + self.eps)
hidden_states = torch.matmul(value, scores.to(value.dtype))
return hidden_states
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.processor(self, hidden_states)
class MochiAttention(nn.Module):
def __init__(
self,
query_dim: int,
added_kv_proj_dim: int,
processor: "MochiAttnProcessor2_0",
heads: int = 8,
dim_head: int = 64,
dropout: float = 0.0,
bias: bool = False,
added_proj_bias: bool = True,
out_dim: int | None = None,
out_context_dim: int | None = None,
out_bias: bool = True,
context_pre_only: bool = False,
eps: float = 1e-5,
):
super().__init__()
from .normalization import MochiRMSNorm
self.inner_dim = out_dim if out_dim is not None else dim_head * heads
self.out_dim = out_dim if out_dim is not None else query_dim
self.out_context_dim = out_context_dim if out_context_dim else query_dim
self.context_pre_only = context_pre_only
self.heads = out_dim // dim_head if out_dim is not None else heads
self.norm_q = MochiRMSNorm(dim_head, eps, True)
self.norm_k = MochiRMSNorm(dim_head, eps, True)
self.norm_added_q = MochiRMSNorm(dim_head, eps, True)
self.norm_added_k = MochiRMSNorm(dim_head, eps, True)
self.to_q = nn.Linear(query_dim, self.inner_dim, bias=bias)
self.to_k = nn.Linear(query_dim, self.inner_dim, bias=bias)
self.to_v = nn.Linear(query_dim, self.inner_dim, bias=bias)
self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
if self.context_pre_only is not None:
self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
self.to_out = nn.ModuleList([])
self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
self.to_out.append(nn.Dropout(dropout))
if not self.context_pre_only:
self.to_add_out = nn.Linear(self.inner_dim, self.out_context_dim, bias=out_bias)
self.processor = processor
def forward(
self,
hidden_states: torch.Tensor,
encoder_hidden_states: torch.Tensor | None = None,
attention_mask: torch.Tensor | None = None,
**kwargs,
) -> tuple[torch.Tensor, torch.Tensor]:
return self.processor(
self,
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
**kwargs,
)
class MochiAttnProcessor2_0:
"""Attention processor used in Mochi."""